IMW 2020: Leaderboard
Current version: 6147c3cf (2021-05-11, 17:43 UTC)
Categories are non-exclusive: e.g. small descriptors compete with larger descriptors, but not vice-versa. For the sake of clarity, we do separate 2k and 8k in the leaderboards. Please note that this is a static website: you may want to force a reload if it does not update properly. We are working on new features and tracks: the organizers reserve the right to update terms and conditions within the next few weeks.
Prize #1: unlimited keypoints (8k) / standard descriptors (512 bytes)
Stereo | Multiview | Combined |
---|---|---|
#1: [sid:00591] Guided-Hardnet-Epoch...
mAA = 0.61081 #2: [sid:00624] Guided-HardNet-OANet
mAA = 0.60261 #3: [sid:00590] Guided-HardNet-epoch...
mAA = 0.59919 |
#1: [sid:00624] Guided-HardNet-OANet
mAA = 0.78550 #2: [sid:00591] Guided-Hardnet-Epoch...
mAA = 0.78288 #3: [sid:00610] Hardnet-Upright-AdaL...
mAA = 0.77056 |
(Total: 144)
#1: [sid:00591] Guided-Hardnet-Epoch...
mAA = 0.69684 #2: [sid:00624] Guided-HardNet-OANet
mAA = 0.69405 #3: [sid:00590] Guided-HardNet-epoch...
mAA = 0.68069 |
Prize #2: restricted keypoints (2k) / standard descriptors (512 bytes)
Stereo | Multiview | Combined |
---|---|---|
#1: [sid:00612] SuperPoint-128d-adap...
mAA = 0.59034 #2: [sid:00667] sp_ae_sg_degensac_xp
mAA = 0.58918 #3: [sid:00603] SuperPoint-128d-mask...
mAA = 0.56769 |
#1: [sid:00667] sp_ae_sg_degensac_xp
mAA = 0.77685 #2: [sid:00612] SuperPoint-128d-adap...
mAA = 0.77337 #3: [sid:00603] SuperPoint-128d-mask...
mAA = 0.76987 |
(Total: 94)
#1: [sid:00667] sp_ae_sg_degensac_xp
mAA = 0.68301 #2: [sid:00612] SuperPoint-128d-adap...
mAA = 0.68186 #3: [sid:00603] SuperPoint-128d-mask...
mAA = 0.66878 |
Phototourism: unlimited keypoints / standard descriptors (512 bytes)
Note: entries with the same multi-view configuration may seem duplicated. This is normal: performance is averaged across tasks.
Stereo | Multiview | Avg. | ||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Method | NF | NI | Rep. (3 pix.) |
MS (3 pix.) |
mAA (at 10o) |
NM | NL | TL | ATE | mAA (at 100) |
mAA (at 100) |
|||||||||
Submission ID: 00081 LogPolar w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 591.21 Rank: 33/144 |
0.472 Rank: 102/144 |
0.832 Rank: 83/144 |
0.52385 (±0.00113) Rank: 57/144 |
415.77 Rank: 90/144 |
4054.60 Rank: 70/144 |
4.316 Rank: 84/144 |
0.432 Rank: 68/144 |
0.69284 (±0.00429) Rank: 70/144 |
0.60835 Rank: 63/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00121 CV-DoG-HardNetAmos-8kSize: 512 bytes. Matches: built-in |
7860.98 | 398.65 Rank: 81/144 |
0.472 Rank: 88/144 |
0.863 Rank: 30/144 |
0.53852 (±0.00039) Rank: 44/144 |
356.56 Rank: 107/144 |
3550.63 Rank: 92/144 |
4.275 Rank: 89/144 |
0.439 Rank: 74/144 |
0.68875 (±0.00138) Rank: 76/144 |
0.61364 Rank: 52/144 |
Challenge organizers (contact) | sift8k | hardnetamos (128 float32: 512 bytes) | CV-DoG-HardNetAmos with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00142 CV-DoG-MKD-Concat-pyransacSize: 512 bytes. Matches: built-in |
7860.77 | 208.02 Rank: 134/144 |
0.472 Rank: 82/144 |
0.854 Rank: 34/144 |
0.40612 (±0.00056) Rank: 118/144 |
348.03 Rank: 119/144 |
3507.39 Rank: 98/144 |
4.169 Rank: 112/144 |
0.471 Rank: 113/144 |
0.64763 (±0.00344) Rank: 108/144 |
0.52687 Rank: 117/144 |
Challenge organizers (contact) | sift | mkd-concat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the MKD-Concat descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1811.11147 | N/A | 21-02-05 | is_baseline | |
Submission ID: 00562 HardNet64-train-all-raw64-balanc...Size: 512 bytes. Matches: built-in |
7831.92 | 370.11 Rank: 91/144 |
0.486 Rank: 77/144 |
0.807 Rank: 114/144 |
0.45770 (±0.00060) Rank: 111/144 |
575.43 Rank: 44/144 |
4756.46 Rank: 33/144 |
4.563 Rank: 36/144 |
0.435 Rank: 72/144 |
0.69780 (±0.00191) Rank: 65/144 |
0.57775 Rank: 89/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-raw64-balance-l2-224000-raw64 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 224000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00602 retrained sosnetSize: 512 bytes. Matches: built-in |
7861.11 | 340.97 Rank: 104/144 |
0.472 Rank: 102/144 |
0.840 Rank: 70/144 |
0.48990 (±0.00046) Rank: 88/144 |
623.55 Rank: 32/144 |
4812.36 Rank: 26/144 |
4.233 Rank: 101/144 |
0.458 Rank: 99/144 |
0.66502 (±0.00158) Rank: 95/144 |
0.57746 Rank: 90/144 |
sosnet (contact) | siftdef | sosnet (128 float32: 512 bytes) | default setting and retrained sosnet | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00576 Upright-Sift + X-Net-lib w/ DEGE...Size: 512 bytes. Matches: built-in |
7830.16 | 465.31 Rank: 62/144 |
0.486 Rank: 44/144 |
0.844 Rank: 61/144 |
0.53156 (±0.00027) Rank: 50/144 |
720.29 Rank: 25/144 |
5398.84 Rank: 19/144 |
4.547 Rank: 41/144 |
0.408 Rank: 37/144 |
0.71966 (±0.00160) Rank: 41/144 |
0.62561 Rank: 46/144 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | sift | x-net-lib (128 float32: 512 bytes) | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | ||
Submission ID: 00086 LogPolar-Upright w/ MAGSAC (no F...Size: 512 bytes. Matches: built-in |
7829.63 | 732.28 Rank: 15/144 |
0.486 Rank: 45/144 |
0.844 Rank: 63/144 |
0.54060 (±0.00047) Rank: 42/144 |
505.37 Rank: 68/144 |
4414.11 Rank: 52/144 |
4.518 Rank: 46/144 |
0.421 Rank: 52/144 |
0.71092 (±0.00251) Rank: 46/144 |
0.62576 Rank: 45/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00061 L2-Net-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 358.37 Rank: 96/144 |
0.486 Rank: 45/144 |
0.845 Rank: 56/144 |
0.52012 (±0.00106) Rank: 62/144 |
369.60 Rank: 105/144 |
3538.91 Rank: 95/144 |
4.407 Rank: 60/144 |
0.447 Rank: 89/144 |
0.68108 (±0.00177) Rank: 85/144 |
0.60060 Rank: 72/144 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-23 | is_baseline | |
Submission ID: 00084 LogPolar-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 459.75 Rank: 63/144 |
0.486 Rank: 45/144 |
0.852 Rank: 42/144 |
0.52316 (±0.00034) Rank: 59/144 |
483.93 Rank: 70/144 |
4405.57 Rank: 55/144 |
4.542 Rank: 42/144 |
0.422 Rank: 53/144 |
0.70680 (±0.00247) Rank: 53/144 |
0.61498 Rank: 50/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00638 MKDNet-polarSize: 512 bytes. Matches: built-in |
7862.74 | 337.43 Rank: 106/144 |
0.472 Rank: 91/144 |
0.847 Rank: 51/144 |
0.51376 (±0.00106) Rank: 70/144 |
552.91 Rank: 47/144 |
4786.70 Rank: 28/144 |
4.259 Rank: 96/144 |
0.434 Rank: 71/144 |
0.69907 (±0.00048) Rank: 63/144 |
0.60642 Rank: 69/144 |
(contact) | sift | mkdnet-polar (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-11 | is_submission | |
Submission ID: 00140 CV-DoG-TFeat-kornia-MAGSACSize: 512 bytes. Matches: built-in |
7860.77 | 292.56 Rank: 118/144 |
0.472 Rank: 82/144 |
0.819 Rank: 106/144 |
0.46677 (±0.00109) Rank: 102/144 |
265.53 Rank: 137/144 |
2905.25 Rank: 125/144 |
4.038 Rank: 128/144 |
0.487 Rank: 120/144 |
0.62608 (±0.00138) Rank: 119/144 |
0.54643 Rank: 109/144 |
Challenge organizers (contact) | sift | tfeat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the TFeat descriptor. Implementation: OpenCV + kornia library | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | N/A | 21-02-05 | is_baseline | |
Submission ID: 00634 SIFT8k-HardNetPS-first-submitSize: 512 bytes. Matches: built-in |
7830.09 | 502.15 Rank: 54/144 |
0.486 Rank: 31/144 |
0.841 Rank: 66/144 |
0.49113 (±0.00083) Rank: 85/144 |
651.33 Rank: 27/144 |
5100.71 Rank: 21/144 |
4.370 Rank: 73/144 |
0.423 Rank: 56/144 |
0.70782 (±0.00264) Rank: 51/144 |
0.59948 Rank: 73/144 |
Xudong Zhang, Yuhao Zhou, Huanhuan Fan (contact) | sift | hardnetps (128 float32: 512 bytes) | SIFT and hardnet with 8k features for fisrt subimission. | N/A | N/A | 20-12-15 | is_submission | |
Submission ID: 00601 SIFT8k-HardNet64Size: 512 bytes. Matches: built-in |
6589.88 | 381.80 Rank: 85/144 |
0.467 Rank: 123/144 |
0.852 Rank: 39/144 |
0.52493 (±0.00044) Rank: 55/144 |
420.36 Rank: 87/144 |
3586.82 Rank: 91/144 |
4.494 Rank: 51/144 |
0.429 Rank: 64/144 |
0.70580 (±0.00229) Rank: 55/144 |
0.61536 Rank: 49/144 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, hardnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-05-29 | is_submission, is_challenge_2020 | |
Submission ID: 00591 Guided-Hardnet-Epoch2Size: 512 bytes. Matches: custom |
7829.63 | 761.99 Rank: 13/144 |
0.486 Rank: 45/144 |
0.823 Rank: 102/144 |
0.61081 (±0.00000) Rank: 1/144 |
785.57 Rank: 23/144 |
6330.70 Rank: 7/144 |
4.680 Rank: 7/144 |
0.358 Rank: 2/144 |
0.78288 (±0.00094) Rank: 2/144 |
0.69684 Rank: 1/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 8k features, using the modified guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-24 | is_submission, is_challenge_2020 | |
Submission ID: 00594 CV-DoG-HardNet8-PTSize: 512 bytes. Matches: built-in |
7829.26 | 579.51 Rank: 37/144 |
0.484 Rank: 78/144 |
0.872 Rank: 15/144 |
0.58375 (±0.00046) Rank: 11/144 |
573.22 Rank: 45/144 |
4489.23 Rank: 47/144 |
4.619 Rank: 18/144 |
0.403 Rank: 31/144 |
0.72432 (±0.00357) Rank: 34/144 |
0.65403 Rank: 22/144 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift8k | h8e512pt (128 float32: 512 bytes) | HardNet8 with PCA compression | N/A | N/A | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00611 sift and hardnet64 train scale(1...Size: 512 bytes. Matches: built-in |
7830.09 | 622.13 Rank: 22/144 |
0.486 Rank: 33/144 |
0.871 Rank: 17/144 |
0.58870 (±0.00041) Rank: 5/144 |
899.14 Rank: 14/144 |
6086.16 Rank: 12/144 |
4.647 Rank: 10/144 |
0.386 Rank: 15/144 |
0.75127 (±0.00234) Rank: 14/144 |
0.66999 Rank: 8/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-l2-val-14000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 14000), FLANN disabled | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00597 SIFT8k-HardNet64Size: 512 bytes. Matches: built-in |
6589.88 | 381.75 Rank: 86/144 |
0.467 Rank: 123/144 |
0.852 Rank: 40/144 |
0.52511 (±0.00056) Rank: 54/144 |
420.42 Rank: 86/144 |
3595.93 Rank: 90/144 |
4.496 Rank: 50/144 |
0.417 Rank: 45/144 |
0.70738 (±0.00218) Rank: 52/144 |
0.61625 Rank: 47/144 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, hardnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00636 MKDNet-cartSize: 512 bytes. Matches: built-in |
7862.74 | 260.22 Rank: 126/144 |
0.472 Rank: 91/144 |
0.819 Rank: 107/144 |
0.46478 (±0.00044) Rank: 106/144 |
432.81 Rank: 83/144 |
4142.25 Rank: 68/144 |
4.050 Rank: 127/144 |
0.463 Rank: 104/144 |
0.66048 (±0.00254) Rank: 98/144 |
0.56263 Rank: 102/144 |
(contact) | sift | mkdnet-cart (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-13 | is_submission | |
Submission ID: 00050 GeoDesc-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 410.23 Rank: 76/144 |
0.486 Rank: 45/144 |
0.827 Rank: 95/144 |
0.48259 (±0.00060) Rank: 94/144 |
394.67 Rank: 102/144 |
3863.42 Rank: 79/144 |
4.380 Rank: 67/144 |
0.443 Rank: 81/144 |
0.67682 (±0.00127) Rank: 90/144 |
0.57971 Rank: 87/144 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00076 SOSNet w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 563.27 Rank: 41/144 |
0.472 Rank: 102/144 |
0.846 Rank: 54/144 |
0.55170 (±0.00042) Rank: 32/144 |
508.66 Rank: 65/144 |
4502.19 Rank: 45/144 |
4.377 Rank: 71/144 |
0.405 Rank: 35/144 |
0.71820 (±0.00298) Rank: 42/144 |
0.63495 Rank: 35/144 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00511 r2d2-wasfiSize: 512 bytes. Matches: built-in |
7861.03 | 276.55 Rank: 121/144 |
0.503 Rank: 17/144 |
0.762 Rank: 123/144 |
0.37893 (±0.00023) Rank: 125/144 |
416.35 Rank: 89/144 |
2834.69 Rank: 128/144 |
4.293 Rank: 88/144 |
0.480 Rank: 117/144 |
0.64359 (±0.00103) Rank: 111/144 |
0.51126 Rank: 124/144 |
Chen Shen (contact) | r2d2 | r2d2-wasfi-epoch25-pretrained-n16-8k (128 float32: 512 bytes) | r2d2 with 8k features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00067 GeoDesc w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 453.39 Rank: 65/144 |
0.472 Rank: 102/144 |
0.835 Rank: 79/144 |
0.50559 (±0.00081) Rank: 77/144 |
395.12 Rank: 100/144 |
3838.97 Rank: 81/144 |
4.264 Rank: 94/144 |
0.442 Rank: 79/144 |
0.68032 (±0.00099) Rank: 88/144 |
0.59295 Rank: 82/144 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00079 LogPolar w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 441.77 Rank: 68/144 |
0.472 Rank: 102/144 |
0.852 Rank: 38/144 |
0.53396 (±0.00086) Rank: 48/144 |
415.77 Rank: 90/144 |
4054.60 Rank: 70/144 |
4.316 Rank: 84/144 |
0.432 Rank: 70/144 |
0.69284 (±0.00429) Rank: 70/144 |
0.61340 Rank: 53/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00508 modified-r2d2Size: 512 bytes. Matches: built-in |
7791.82 | 160.64 Rank: 141/144 |
0.517 Rank: 15/144 |
0.785 Rank: 119/144 |
0.32815 (±0.00037) Rank: 127/144 |
328.39 Rank: 128/144 |
2811.27 Rank: 129/144 |
4.250 Rank: 98/144 |
0.488 Rank: 121/144 |
0.62899 (±0.00193) Rank: 115/144 |
0.47857 Rank: 127/144 |
caoliang (contact) | modified-r2d2 | modified-r2d2 (128 float32: 512 bytes) | r2d2 with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | https://europe.naverlabs.com/wp-content/uploads/2019/09/R2D2-Repeatable-and-Reliable-Detector-and-Descriptor-2.pdf | https://github.com/naver/r2d2 | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00514 UprightRootSIFT-AdaLAMSize: 512 bytes. Matches: custom |
6449.42 | 619.87 Rank: 23/144 |
0.436 Rank: 131/144 |
0.758 Rank: 125/144 |
0.31372 (±0.00000) Rank: 129/144 |
640.08 Rank: 30/144 |
5286.27 Rank: 20/144 |
4.343 Rank: 76/144 |
0.398 Rank: 25/144 |
0.72130 (±0.00085) Rank: 40/144 |
0.51751 Rank: 119/144 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Using upright RootSIFT with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. | https://arxiv.org/abs/2006.04250 | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00052 HardNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 363.51 Rank: 93/144 |
0.472 Rank: 102/144 |
0.853 Rank: 37/144 |
0.52325 (±0.00098) Rank: 58/144 |
402.15 Rank: 93/144 |
3895.00 Rank: 77/144 |
4.345 Rank: 74/144 |
0.426 Rank: 61/144 |
0.69653 (±0.00108) Rank: 66/144 |
0.60989 Rank: 60/144 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00501 Guided matching with Upright roo...Size: 512 bytes. Matches: custom |
7829.27 | 609.35 Rank: 27/144 |
0.486 Rank: 29/144 |
0.786 Rank: 118/144 |
0.49072 (±0.00000) Rank: 86/144 |
1363.98 Rank: 3/144 |
7835.75 Rank: 1/144 |
4.219 Rank: 103/144 |
0.454 Rank: 95/144 |
0.68064 (±0.00390) Rank: 87/144 |
0.58568 Rank: 85/144 |
(contact) | sift | upright-root-sift (128 float32: 512 bytes) | In submission | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00122 CV-DoG-HardNetAmos-8kSize: 512 bytes. Matches: built-in |
7860.98 | 528.67 Rank: 45/144 |
0.472 Rank: 88/144 |
0.838 Rank: 72/144 |
0.53291 (±0.00047) Rank: 49/144 |
356.56 Rank: 107/144 |
3550.63 Rank: 92/144 |
4.275 Rank: 89/144 |
0.443 Rank: 82/144 |
0.68875 (±0.00138) Rank: 76/144 |
0.61083 Rank: 58/144 |
Challenge organizers (contact) | sift8k | hardnetamos (128 float32: 512 bytes) | CV-DoG-HardNetAmos with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00558 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 333.94 Rank: 108/144 |
0.284 Rank: 140/144 |
0.427 Rank: 139/144 |
0.01806 (±0.00000) Rank: 142/144 |
341.75 Rank: 125/144 |
2003.56 Rank: 136/144 |
3.735 Rank: 135/144 |
0.593 Rank: 134/144 |
0.51048 (±0.00436) Rank: 134/144 |
0.26427 Rank: 138/144 |
Fabio Bellavia (contact) | sift | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 10) [Mishkin 2019] | N/A | N/A | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00614 ContextDesc Upright + Mutual Che...Size: 512 bytes. Matches: custom |
7830.09 | 647.63 Rank: 18/144 |
0.487 Rank: 25/144 |
0.846 Rank: 55/144 |
0.57826 (±0.00000) Rank: 17/144 |
668.00 Rank: 26/144 |
5612.57 Rank: 16/144 |
4.666 Rank: 8/144 |
0.367 Rank: 4/144 |
0.77041 (±0.00298) Rank: 5/144 |
0.67433 Rank: 5/144 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | contextdesc-upright | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00625 HardNet64-train-all-l2-val-14000...Size: 512 bytes. Matches: built-in |
7830.09 | 605.12 Rank: 29/144 |
0.486 Rank: 33/144 |
0.872 Rank: 16/144 |
0.58776 (±0.00048) Rank: 7/144 |
899.14 Rank: 14/144 |
6095.78 Rank: 11/144 |
4.645 Rank: 11/144 |
0.389 Rank: 19/144 |
0.74849 (±0.00117) Rank: 17/144 |
0.66812 Rank: 11/144 |
Ximin Zheng, Sheng He, Guanlin Liang (contact) | sift8k | hardnet64-train-all-l2-val-14000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 14000), FLANN disabled | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00540 ASLV2+OANetV2+DEGENSACSize: 512 bytes. Matches: custom |
5982.87 | 1044.40 Rank: 5/144 |
0.550 Rank: 11/144 |
0.758 Rank: 124/144 |
0.50173 (±0.00000) Rank: 80/144 |
1067.11 Rank: 6/144 |
4587.47 Rank: 43/144 |
4.948 Rank: 2/144 |
0.386 Rank: 16/144 |
0.75275 (±0.00186) Rank: 13/144 |
0.62724 Rank: 43/144 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | aslv2 | aslv2 (128 float32: 512 bytes) | ASL detector and descriptor, 8k keypoints, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00049 GeoDesc-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 317.71 Rank: 113/144 |
0.486 Rank: 45/144 |
0.850 Rank: 46/144 |
0.48769 (±0.00064) Rank: 90/144 |
394.67 Rank: 102/144 |
3863.42 Rank: 79/144 |
4.380 Rank: 67/144 |
0.440 Rank: 76/144 |
0.67682 (±0.00127) Rank: 90/144 |
0.58225 Rank: 86/144 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00600 ContextDesc Upright + Mutual Che...Size: 512 bytes. Matches: custom |
7830.09 | 624.55 Rank: 20/144 |
0.487 Rank: 25/144 |
0.847 Rank: 53/144 |
0.57344 (±0.00000) Rank: 19/144 |
644.39 Rank: 29/144 |
5427.23 Rank: 18/144 |
4.700 Rank: 4/144 |
0.368 Rank: 5/144 |
0.77043 (±0.00133) Rank: 4/144 |
0.67194 Rank: 6/144 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | contextdesc-upright | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-29 | is_submission, is_challenge_2020 | |
Submission ID: 00532 affnet-hardnetSize: 512 bytes. Matches: built-in |
7925.53 | 226.75 Rank: 131/144 |
0.452 Rank: 129/144 |
0.649 Rank: 133/144 |
0.31958 (±0.00055) Rank: 128/144 |
243.28 Rank: 140/144 |
3155.72 Rank: 117/144 |
3.595 Rank: 137/144 |
0.554 Rank: 133/144 |
0.52578 (±0.00343) Rank: 133/144 |
0.42268 Rank: 130/144 |
caoliang (contact) | affnet | hardnet (128 float32: 512 bytes) | affnet up to 8000 keypoints, harnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1711.06704 | https://github.com/ducha-aiki/affnet | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00071 L2-Net w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 366.00 Rank: 92/144 |
0.472 Rank: 102/144 |
0.852 Rank: 44/144 |
0.52953 (±0.00043) Rank: 51/144 |
339.00 Rank: 126/144 |
3424.94 Rank: 103/144 |
4.206 Rank: 105/144 |
0.452 Rank: 93/144 |
0.66445 (±0.00213) Rank: 96/144 |
0.59699 Rank: 77/144 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-23 | is_baseline | |
Submission ID: 00074 L2-Net w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 481.03 Rank: 57/144 |
0.472 Rank: 102/144 |
0.830 Rank: 91/144 |
0.52519 (±0.00039) Rank: 53/144 |
339.00 Rank: 126/144 |
3424.94 Rank: 103/144 |
4.206 Rank: 105/144 |
0.448 Rank: 90/144 |
0.66445 (±0.00213) Rank: 96/144 |
0.59482 Rank: 81/144 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-23 | is_baseline | |
Submission ID: 00054 HardNet w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 477.64 Rank: 58/144 |
0.472 Rank: 102/144 |
0.830 Rank: 90/144 |
0.51809 (±0.00053) Rank: 64/144 |
402.15 Rank: 93/144 |
3895.00 Rank: 77/144 |
4.345 Rank: 74/144 |
0.423 Rank: 58/144 |
0.69653 (±0.00108) Rank: 66/144 |
0.60731 Rank: 67/144 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00073 L2-Net-Upright w/ MAGSAC (no FLA...Size: 512 bytes. Matches: built-in |
7829.63 | 577.91 Rank: 38/144 |
0.486 Rank: 45/144 |
0.837 Rank: 78/144 |
0.53911 (±0.00062) Rank: 43/144 |
395.53 Rank: 98/144 |
3603.85 Rank: 86/144 |
4.382 Rank: 64/144 |
0.455 Rank: 96/144 |
0.68491 (±0.00338) Rank: 81/144 |
0.61201 Rank: 57/144 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00572 Guided HardNet qhtSize: 512 bytes. Matches: custom |
7829.63 | 885.09 Rank: 8/144 |
0.486 Rank: 45/144 |
0.774 Rank: 122/144 |
0.54663 (±0.00000) Rank: 35/144 |
538.28 Rank: 52/144 |
4687.01 Rank: 35/144 |
4.644 Rank: 13/144 |
0.376 Rank: 10/144 |
0.75721 (±0.00140) Rank: 11/144 |
0.65192 Rank: 25/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-qht (128 float32: 512 bytes) | sift and hardnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | |
Submission ID: 00621 CV-DoG-HardNet8-PTv2Size: 512 bytes. Matches: built-in |
7830.09 | 583.12 Rank: 35/144 |
0.486 Rank: 31/144 |
0.880 Rank: 4/144 |
0.58997 (±0.00054) Rank: 4/144 |
577.37 Rank: 42/144 |
4476.61 Rank: 48/144 |
4.638 Rank: 15/144 |
0.407 Rank: 36/144 |
0.73096 (±0.00227) Rank: 24/144 |
0.66046 Rank: 17/144 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift8k | h8e512pt (128 float32: 512 bytes) | HardNet8 with PCA compression, batch sampling from few images | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00138 CV-DoG-TFeat-kornia-DEGENSACSize: 512 bytes. Matches: built-in |
7860.77 | 234.75 Rank: 130/144 |
0.472 Rank: 82/144 |
0.831 Rank: 86/144 |
0.46486 (±0.00039) Rank: 105/144 |
265.53 Rank: 137/144 |
2905.25 Rank: 125/144 |
4.038 Rank: 128/144 |
0.487 Rank: 119/144 |
0.62608 (±0.00138) Rank: 119/144 |
0.54547 Rank: 111/144 |
Challenge organizers (contact) | sift | tfeat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the TFeat descriptor. Implementation: OpenCV + kornia library | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | N/A | 21-02-05 | is_baseline | |
Submission ID: 00065 SOSNet-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 407.29 Rank: 78/144 |
0.486 Rank: 45/144 |
0.864 Rank: 29/144 |
0.54098 (±0.00086) Rank: 41/144 |
521.37 Rank: 60/144 |
4473.48 Rank: 49/144 |
4.597 Rank: 23/144 |
0.417 Rank: 43/144 |
0.72179 (±0.00330) Rank: 37/144 |
0.63139 Rank: 38/144 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00595 SIFT8k_8000_HardNet64-train-all-...Size: 512 bytes. Matches: built-in |
7830.09 | 600.42 Rank: 31/144 |
0.486 Rank: 33/144 |
0.867 Rank: 23/144 |
0.58265 (±0.00102) Rank: 14/144 |
786.40 Rank: 22/144 |
5595.19 Rank: 17/144 |
4.603 Rank: 22/144 |
0.390 Rank: 20/144 |
0.74372 (±0.00317) Rank: 23/144 |
0.66319 Rank: 16/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-ben86-l2-210000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 210000 and iou 0.86), FLANN disabled | N/A | N/A | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00566 Guided-HardNet2qhtp.jsonSize: 512 bytes. Matches: custom |
7829.63 | 517.72 Rank: 49/144 |
0.486 Rank: 45/144 |
0.874 Rank: 11/144 |
0.58251 (±0.00000) Rank: 15/144 |
448.12 Rank: 79/144 |
4072.82 Rank: 69/144 |
4.607 Rank: 20/144 |
0.387 Rank: 17/144 |
0.74640 (±0.00246) Rank: 20/144 |
0.66445 Rank: 14/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnetnd32qhtp (128 float32: 512 bytes) | sift and hardnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-15 | is_submission, is_challenge_2020 | |
Submission ID: 00080 LogPolar w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 376.42 Rank: 88/144 |
0.472 Rank: 102/144 |
0.837 Rank: 76/144 |
0.50188 (±0.00044) Rank: 79/144 |
399.70 Rank: 95/144 |
4036.37 Rank: 72/144 |
4.340 Rank: 77/144 |
0.429 Rank: 65/144 |
0.69100 (±0.00154) Rank: 73/144 |
0.59644 Rank: 78/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00019 Upright Root-SIFT (OpenCV), DEGE...Size: 512 bytes. Matches: built-in |
7829.24 | 355.69 Rank: 97/144 |
0.487 Rank: 23/144 |
0.844 Rank: 64/144 |
0.51009 (±0.00033) Rank: 74/144 |
548.18 Rank: 49/144 |
4406.41 Rank: 54/144 |
4.385 Rank: 62/144 |
0.442 Rank: 78/144 |
0.68859 (±0.00254) Rank: 79/144 |
0.59934 Rank: 74/144 |
Challenge organizers (contact) | sift-lowth | rootsift-upright (128 float32: 512 bytes) | Upright Root-SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00062 L2-Net-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 471.61 Rank: 60/144 |
0.486 Rank: 45/144 |
0.822 Rank: 103/144 |
0.51450 (±0.00058) Rank: 68/144 |
369.60 Rank: 105/144 |
3538.91 Rank: 95/144 |
4.407 Rank: 60/144 |
0.448 Rank: 91/144 |
0.68108 (±0.00177) Rank: 85/144 |
0.59779 Rank: 76/144 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00628 Multiple kernel local descriptorSize: 512 bytes. Matches: built-in |
7862.74 | 331.20 Rank: 109/144 |
0.472 Rank: 91/144 |
0.829 Rank: 92/144 |
0.48380 (±0.00048) Rank: 93/144 |
541.02 Rank: 51/144 |
4639.37 Rank: 40/144 |
4.144 Rank: 116/144 |
0.463 Rank: 105/144 |
0.66704 (±0.00645) Rank: 93/144 |
0.57542 Rank: 92/144 |
(contact) | sift | mkdpcawt (128 float32: 512 bytes) | based on the paper [Understanding and Improving Kernel Local Descriptors] | https://arxiv.org/pdf/1811.11147 | https://arxiv.org/abs/1811.11147 | 20-10-28 | is_submission | |
Submission ID: 00522 SIFT and HardNet64 train scale(5...Size: 512 bytes. Matches: built-in |
7830.09 | 513.57 Rank: 51/144 |
0.486 Rank: 33/144 |
0.852 Rank: 45/144 |
0.55385 (±0.00076) Rank: 30/144 |
822.07 Rank: 18/144 |
5660.19 Rank: 15/144 |
4.555 Rank: 37/144 |
0.404 Rank: 33/144 |
0.72744 (±0.00455) Rank: 28/144 |
0.64065 Rank: 31/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-raw64 (128 float32: 512 bytes) | SIFT with 8000 keypoints(raw 64), hardnet64 with 128 descriptors, FLANN disabled | N/A | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00132 CV-DoG-AffNet-HardNet-kornia-MAG...Size: 512 bytes. Matches: built-in |
7833.97 | 516.85 Rank: 50/144 |
0.486 Rank: 74/144 |
0.870 Rank: 19/144 |
0.54116 (±0.00028) Rank: 40/144 |
580.47 Rank: 38/144 |
4671.35 Rank: 37/144 |
4.565 Rank: 33/144 |
0.403 Rank: 27/144 |
0.72668 (±0.00115) Rank: 30/144 |
0.63392 Rank: 36/144 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00057 HardNet-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 582.52 Rank: 36/144 |
0.486 Rank: 45/144 |
0.841 Rank: 67/144 |
0.53730 (±0.00082) Rank: 46/144 |
480.13 Rank: 72/144 |
4167.34 Rank: 63/144 |
4.566 Rank: 31/144 |
0.416 Rank: 42/144 |
0.71542 (±0.00274) Rank: 44/144 |
0.62636 Rank: 44/144 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00048 GeoDesc w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 362.25 Rank: 94/144 |
0.472 Rank: 102/144 |
0.816 Rank: 110/144 |
0.46887 (±0.00025) Rank: 99/144 |
350.43 Rank: 117/144 |
3601.67 Rank: 88/144 |
4.234 Rank: 99/144 |
0.447 Rank: 88/144 |
0.65553 (±0.00444) Rank: 102/144 |
0.56220 Rank: 103/144 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00120 CV-DoG-HardNetAmos-8kSize: 512 bytes. Matches: built-in |
7860.98 | 265.75 Rank: 124/144 |
0.472 Rank: 88/144 |
0.873 Rank: 13/144 |
0.46073 (±0.00062) Rank: 110/144 |
356.56 Rank: 107/144 |
3550.63 Rank: 92/144 |
4.275 Rank: 89/144 |
0.439 Rank: 75/144 |
0.68875 (±0.00138) Rank: 76/144 |
0.57474 Rank: 93/144 |
Challenge organizers (contact) | sift8k | hardnetamos (128 float32: 512 bytes) | CV-DoG-HardNetAmos with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00502 Guided matching with Upright roo...Size: 512 bytes. Matches: custom |
7829.27 | 491.96 Rank: 56/144 |
0.486 Rank: 29/144 |
0.828 Rank: 94/144 |
0.50310 (±0.01761) Rank: 78/144 |
809.76 Rank: 20/144 |
5966.91 Rank: 13/144 |
4.462 Rank: 54/144 |
0.425 Rank: 60/144 |
0.71009 (±0.00414) Rank: 48/144 |
0.60659 Rank: 68/144 |
(contact) | sift | upright-root-sift (128 float32: 512 bytes) | In submission | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00510 sift and hardnet64Size: 512 bytes. Matches: built-in |
7861.62 | 326.87 Rank: 110/144 |
0.472 Rank: 100/144 |
0.845 Rank: 59/144 |
0.51149 (±0.00018) Rank: 72/144 |
521.58 Rank: 59/144 |
4587.88 Rank: 42/144 |
4.222 Rank: 102/144 |
0.424 Rank: 59/144 |
0.69458 (±0.00340) Rank: 68/144 |
0.60304 Rank: 70/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-v2 (128 float32: 512 bytes) | SIFT with 8000 keypoints, hardnet64 with 128 descriptors | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00106 VL-DoG-SIFT-8kSize: 512 bytes. Matches: built-in |
7880.59 | 326.20 Rank: 111/144 |
0.490 Rank: 20/144 |
0.809 Rank: 113/144 |
0.46326 (±0.00089) Rank: 107/144 |
324.62 Rank: 129/144 |
3030.67 Rank: 121/144 |
4.173 Rank: 109/144 |
0.462 Rank: 103/144 |
0.62829 (±0.00112) Rank: 116/144 |
0.54578 Rank: 110/144 |
Challenge organizers (contact) | dog | vlsift (128 float32: 512 bytes) | VL-DoG-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00618 Sift-Sem + HyNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
7530.17 | 476.84 Rank: 59/144 |
0.490 Rank: 19/144 |
0.882 Rank: 3/144 |
0.57281 (±0.00047) Rank: 21/144 |
809.82 Rank: 19/144 |
5868.33 Rank: 14/144 |
4.585 Rank: 26/144 |
0.390 Rank: 22/144 |
0.74557 (±0.00393) Rank: 21/144 |
0.65919 Rank: 19/144 |
Barroso-Laguna, Axel and Tian, Yurun, Ng, Tony and Mikolajczyk, Krystian (contact) | sift-semantics | hynet (128 float32: 512 bytes) | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | ||
Submission ID: 00068 GeoDesc-Upright w/ DEGENSAC (no ...Size: 512 bytes. Matches: built-in |
7829.63 | 409.87 Rank: 77/144 |
0.486 Rank: 45/144 |
0.869 Rank: 20/144 |
0.52674 (±0.00049) Rank: 52/144 |
458.60 Rank: 75/144 |
4146.83 Rank: 66/144 |
4.412 Rank: 58/144 |
0.431 Rank: 67/144 |
0.70442 (±0.00189) Rank: 57/144 |
0.61558 Rank: 48/144 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00070 GeoDesc-Upright w/ MAGSAC (no FL...Size: 512 bytes. Matches: built-in |
7829.63 | 534.05 Rank: 44/144 |
0.486 Rank: 45/144 |
0.848 Rank: 49/144 |
0.52078 (±0.00028) Rank: 61/144 |
458.60 Rank: 75/144 |
4146.83 Rank: 66/144 |
4.412 Rank: 58/144 |
0.432 Rank: 69/144 |
0.70442 (±0.00189) Rank: 57/144 |
0.61260 Rank: 56/144 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00537 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 1476.71 Rank: 1/144 |
0.284 Rank: 140/144 |
0.145 Rank: 144/144 |
0.00178 (±0.00000) Rank: 144/144 |
1506.82 Rank: 2/144 |
2538.44 Rank: 132/144 |
3.388 Rank: 141/144 |
0.669 Rank: 139/144 |
0.42682 (±0.00259) Rank: 139/144 |
0.21430 Rank: 141/144 |
Fabio Bellavia (contact) | sift | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00536 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 1164.84 Rank: 3/144 |
0.368 Rank: 135/144 |
0.341 Rank: 141/144 |
0.00607 (±0.00000) Rank: 143/144 |
1191.31 Rank: 4/144 |
2215.65 Rank: 134/144 |
4.014 Rank: 132/144 |
0.524 Rank: 132/144 |
0.58741 (±0.00250) Rank: 128/144 |
0.29674 Rank: 136/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00075 SOSNet w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 424.61 Rank: 74/144 |
0.472 Rank: 102/144 |
0.868 Rank: 22/144 |
0.55867 (±0.00060) Rank: 26/144 |
508.66 Rank: 65/144 |
4502.19 Rank: 45/144 |
4.377 Rank: 71/144 |
0.404 Rank: 32/144 |
0.71820 (±0.00298) Rank: 42/144 |
0.63844 Rank: 32/144 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-23 | is_baseline | |
Submission ID: 00059 L2-Net w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 297.34 Rank: 117/144 |
0.472 Rank: 102/144 |
0.835 Rank: 81/144 |
0.49523 (±0.00042) Rank: 83/144 |
314.54 Rank: 132/144 |
3314.19 Rank: 112/144 |
4.204 Rank: 107/144 |
0.467 Rank: 110/144 |
0.65665 (±0.00287) Rank: 99/144 |
0.57594 Rank: 91/144 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00042 ORB (OpenCV), DEGENSACSize: 32 bytes. Matches: built-in |
7150.21 | 161.98 Rank: 139/144 |
0.514 Rank: 16/144 |
0.653 Rank: 131/144 |
0.16159 (±0.00090) Rank: 134/144 |
910.31 Rank: 13/144 |
1423.38 Rank: 142/144 |
2.722 Rank: 143/144 |
0.897 Rank: 144/144 |
0.08054 (±0.00242) Rank: 144/144 |
0.12106 Rank: 144/144 |
Challenge organizers (contact) | orb | orb (32 uint8: 32 bytes) | ORB with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00520 SIFT and HardNet64 train scale(5...Size: 512 bytes. Matches: built-in |
7830.09 | 400.19 Rank: 80/144 |
0.486 Rank: 33/144 |
0.837 Rank: 77/144 |
0.50589 (±0.00059) Rank: 76/144 |
619.25 Rank: 33/144 |
4959.78 Rank: 24/144 |
4.384 Rank: 63/144 |
0.456 Rank: 97/144 |
0.69172 (±0.00048) Rank: 72/144 |
0.59880 Rank: 75/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-scale5 (128 float32: 512 bytes) | SIFT with 8000 keypoints(size scaled by 5), hardnet64 with 128 descriptors, FLANN disabled | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00055 HardNet-Upright w/ DEGENSAC (no ...Size: 512 bytes. Matches: built-in |
7829.63 | 527.59 Rank: 47/144 |
0.486 Rank: 45/144 |
0.876 Rank: 7/144 |
0.57279 (±0.00120) Rank: 22/144 |
509.07 Rank: 63/144 |
4250.40 Rank: 58/144 |
4.548 Rank: 39/144 |
0.413 Rank: 40/144 |
0.72309 (±0.00141) Rank: 35/144 |
0.64794 Rank: 27/144 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00613 HardNet64-data-aug-sort-51Size: 512 bytes. Matches: built-in |
7830.09 | 624.09 Rank: 21/144 |
0.486 Rank: 33/144 |
0.870 Rank: 18/144 |
0.58727 (±0.00076) Rank: 8/144 |
964.80 Rank: 9/144 |
6350.68 Rank: 5/144 |
4.644 Rank: 14/144 |
0.383 Rank: 13/144 |
0.74952 (±0.00162) Rank: 16/144 |
0.66839 Rank: 10/144 |
Ximin Zheng, Sheng He, Guanlin Liang (contact) | sift8k | hardnet64-data-aug-sort-51 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 124000 and data augument), FLANN disabled | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00554 ContextDesc-Upright w/ DEGENSAC ...Size: 512 bytes. Matches: built-in |
7830.09 | 470.69 Rank: 61/144 |
0.487 Rank: 25/144 |
0.877 Rank: 5/144 |
0.55697 (±0.00079) Rank: 28/144 |
543.45 Rank: 50/144 |
4543.76 Rank: 44/144 |
4.498 Rank: 49/144 |
0.405 Rank: 34/144 |
0.72902 (±0.00255) Rank: 27/144 |
0.64299 Rank: 30/144 |
Zixin Luo, Jiahui Zhang, Hongkai Chen (contact) | sift-def | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value and use only single orientation to increase performance. FLANN disabled. | https://arxiv.org/abs/1904.04084 | https://github.com/lzx551402/contextdesc | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00571 ContextDesc Upright + OANetV2 + ...Size: 512 bytes. Matches: custom |
7830.09 | 827.45 Rank: 9/144 |
0.487 Rank: 25/144 |
0.775 Rank: 120/144 |
0.53810 (±0.00000) Rank: 45/144 |
855.38 Rank: 16/144 |
6561.52 Rank: 3/144 |
4.522 Rank: 44/144 |
0.370 Rank: 8/144 |
0.76702 (±0.00081) Rank: 6/144 |
0.65256 Rank: 23/144 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | contextdesc-upright | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-21 | is_submission, is_challenge_2020 | |
Submission ID: 00078 SOSNet-Upright w/ MAGSAC (no FLA...Size: 512 bytes. Matches: built-in |
7829.63 | 679.94 Rank: 17/144 |
0.486 Rank: 45/144 |
0.855 Rank: 33/144 |
0.56626 (±0.00020) Rank: 23/144 |
600.23 Rank: 35/144 |
4765.43 Rank: 31/144 |
4.584 Rank: 27/144 |
0.403 Rank: 28/144 |
0.73027 (±0.00284) Rank: 25/144 |
0.64826 Rank: 26/144 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00630 Superpoint_modifiedSize: 512 bytes. Matches: built-in |
7927.15 | 25.26 Rank: 144/144 |
0.410 Rank: 134/144 |
0.328 Rank: 142/144 |
0.06569 (±0.00036) Rank: 137/144 |
70.01 Rank: 144/144 |
895.58 Rank: 144/144 |
2.963 Rank: 142/144 |
0.849 Rank: 143/144 |
0.18293 (±0.00230) Rank: 142/144 |
0.12431 Rank: 143/144 |
Anonymous (to be released: 2020-6-12) | superpoint-modified | superpoint-modified (128 float32: 512 bytes) | Modified superpoint output. | N/A | N/A | 20-11-27 | is_submission | |
Submission ID: 00063 SOSNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 340.09 Rank: 105/144 |
0.472 Rank: 102/144 |
0.854 Rank: 35/144 |
0.51891 (±0.00067) Rank: 63/144 |
440.21 Rank: 80/144 |
4178.82 Rank: 61/144 |
4.378 Rank: 69/144 |
0.419 Rank: 48/144 |
0.70078 (±0.00375) Rank: 60/144 |
0.60984 Rank: 61/144 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00548 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 343.78 Rank: 102/144 |
0.368 Rank: 135/144 |
0.607 Rank: 135/144 |
0.28795 (±0.00000) Rank: 131/144 |
351.53 Rank: 113/144 |
1699.08 Rank: 139/144 |
4.124 Rank: 120/144 |
0.507 Rank: 126/144 |
0.61370 (±0.00246) Rank: 122/144 |
0.45082 Rank: 129/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.75, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-03 | is_submission, is_challenge_2020 | |
Submission ID: 00521 SIFT8k-giftSize: 512 bytes. Matches: built-in |
6589.48 | 426.77 Rank: 73/144 |
0.467 Rank: 126/144 |
0.830 Rank: 87/144 |
0.48128 (±0.00056) Rank: 95/144 |
452.18 Rank: 77/144 |
3770.17 Rank: 85/144 |
4.595 Rank: 25/144 |
0.444 Rank: 84/144 |
0.67443 (±0.00253) Rank: 92/144 |
0.57785 Rank: 88/144 |
Chen Shen (contact) | sift-def | gift (128 float32: 512 bytes) | sift-nodups, gift with 8k features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00504 ASLFeat (MS) w/ DEGENSAC (no FLA...Size: 512 bytes. Matches: built-in |
6948.83 | 390.79 Rank: 82/144 |
0.550 Rank: 10/144 |
0.797 Rank: 116/144 |
0.46102 (±0.00067) Rank: 109/144 |
510.80 Rank: 62/144 |
3817.14 Rank: 83/144 |
4.577 Rank: 29/144 |
0.444 Rank: 83/144 |
0.68246 (±0.00173) Rank: 84/144 |
0.57174 Rank: 97/144 |
Zixin Luo, Jiahui Zhang (contact) | aslfeat | aslfeat (128 float32: 512 bytes) | ASLFeat (MS) with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. FLANN disabled. | https://arxiv.org/abs/2003.10071 | https://github.com/lzx551402/ASLFeat | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00010 AKAZE (OpenCV), DEGENSACSize: 61 bytes. Matches: built-in |
7857.11 | 246.74 Rank: 128/144 |
0.553 Rank: 9/144 |
0.735 Rank: 127/144 |
0.30717 (±0.00122) Rank: 130/144 |
479.55 Rank: 74/144 |
2778.68 Rank: 130/144 |
3.393 Rank: 140/144 |
0.737 Rank: 142/144 |
0.36048 (±0.00382) Rank: 141/144 |
0.33383 Rank: 132/144 |
Challenge organizers (contact) | akaze-lowth | akaze (61 uint8: 61 bytes) | AKAZE with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00141 CV-DoG-MKD-Concat-DEGENSACSize: 512 bytes. Matches: built-in |
7860.77 | 305.75 Rank: 115/144 |
0.472 Rank: 82/144 |
0.845 Rank: 58/144 |
0.48465 (±0.00122) Rank: 92/144 |
348.03 Rank: 119/144 |
3507.39 Rank: 98/144 |
4.169 Rank: 112/144 |
0.467 Rank: 109/144 |
0.64763 (±0.00344) Rank: 108/144 |
0.56614 Rank: 99/144 |
Challenge organizers (contact) | sift | mkd-concat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the MKD-Concat descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1811.11147 | N/A | 21-02-05 | is_baseline | |
Submission ID: 00051 HardNet w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 432.32 Rank: 72/144 |
0.472 Rank: 102/144 |
0.866 Rank: 25/144 |
0.55430 (±0.00031) Rank: 29/144 |
426.84 Rank: 84/144 |
4001.40 Rank: 75/144 |
4.339 Rank: 79/144 |
0.419 Rank: 47/144 |
0.70962 (±0.00191) Rank: 49/144 |
0.63196 Rank: 37/144 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00639 MKDNet-hardnetSize: 512 bytes. Matches: built-in |
7862.74 | 341.96 Rank: 103/144 |
0.472 Rank: 91/144 |
0.847 Rank: 52/144 |
0.51643 (±0.00099) Rank: 65/144 |
558.95 Rank: 46/144 |
4811.45 Rank: 27/144 |
4.268 Rank: 92/144 |
0.420 Rank: 51/144 |
0.69845 (±0.00564) Rank: 64/144 |
0.60744 Rank: 65/144 |
(contact) | sift | mkdnet-hardnet (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-11 | is_submission | |
Submission ID: 00005 SURF (OpenCV), DEGENSACSize: 256 bytes. Matches: built-in |
7728.57 | 125.94 Rank: 142/144 |
0.432 Rank: 133/144 |
0.684 Rank: 130/144 |
0.24948 (±0.00065) Rank: 132/144 |
928.29 Rank: 12/144 |
3455.28 Rank: 102/144 |
3.428 Rank: 138/144 |
0.733 Rank: 141/144 |
0.40490 (±0.00448) Rank: 140/144 |
0.32719 Rank: 133/144 |
Challenge organizers (contact) | surf-lowth | surf (64 float32: 256 bytes) | SURF with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00069 GeoDesc w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 348.52 Rank: 101/144 |
0.472 Rank: 102/144 |
0.856 Rank: 32/144 |
0.51112 (±0.00070) Rank: 73/144 |
395.12 Rank: 100/144 |
3838.97 Rank: 81/144 |
4.264 Rank: 94/144 |
0.443 Rank: 80/144 |
0.68032 (±0.00099) Rank: 88/144 |
0.59572 Rank: 80/144 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00559 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 164.43 Rank: 138/144 |
0.284 Rank: 140/144 |
0.650 Rank: 132/144 |
0.05206 (±0.00000) Rank: 139/144 |
167.59 Rank: 142/144 |
1549.91 Rank: 140/144 |
3.989 Rank: 134/144 |
0.622 Rank: 136/144 |
0.47700 (±0.00167) Rank: 136/144 |
0.26453 Rank: 137/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 5) [Mishkin 2019] | N/A | N/A | 20-05-10 | is_submission, is_challenge_2020 | |
Submission ID: 00006 Upright SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
7829.24 | 319.29 Rank: 112/144 |
0.487 Rank: 23/144 |
0.830 Rank: 89/144 |
0.48742 (±0.00042) Rank: 91/144 |
525.57 Rank: 56/144 |
4147.20 Rank: 65/144 |
4.258 Rank: 97/144 |
0.460 Rank: 100/144 |
0.65664 (±0.00231) Rank: 101/144 |
0.57203 Rank: 96/144 |
Challenge organizers (contact) | sift-lowth | sift-upright (128 float32: 512 bytes) | Upright SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00573 Guided-HardNet-qhtSize: 512 bytes. Matches: custom |
7829.63 | 887.83 Rank: 7/144 |
0.486 Rank: 45/144 |
0.774 Rank: 121/144 |
0.55224 (±0.00000) Rank: 31/144 |
538.28 Rank: 52/144 |
4693.85 Rank: 34/144 |
4.650 Rank: 9/144 |
0.381 Rank: 11/144 |
0.75930 (±0.00177) Rank: 8/144 |
0.65577 Rank: 20/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-qht (128 float32: 512 bytes) | sift and hardnet with 8k features, using the modified guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | |
Submission ID: 00066 SOSNet-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 537.85 Rank: 43/144 |
0.486 Rank: 45/144 |
0.841 Rank: 68/144 |
0.53531 (±0.00061) Rank: 47/144 |
521.37 Rank: 60/144 |
4473.48 Rank: 49/144 |
4.597 Rank: 23/144 |
0.414 Rank: 41/144 |
0.72179 (±0.00330) Rank: 37/144 |
0.62855 Rank: 42/144 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00114 VL-Hess-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 348.95 Rank: 100/144 |
0.547 Rank: 12/144 |
0.800 Rank: 115/144 |
0.43347 (±0.00084) Rank: 115/144 |
347.39 Rank: 122/144 |
3209.10 Rank: 114/144 |
4.126 Rank: 117/144 |
0.517 Rank: 128/144 |
0.58657 (±0.00120) Rank: 129/144 |
0.51002 Rank: 125/144 |
Challenge organizers (contact) | hessian | vlsift (128 float32: 512 bytes) | VL-Hess-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00624 Guided-HardNet-OANetSize: 512 bytes. Matches: custom |
7829.63 | 765.34 Rank: 12/144 |
0.486 Rank: 45/144 |
0.820 Rank: 104/144 |
0.60261 (±0.00000) Rank: 2/144 |
788.49 Rank: 21/144 |
6346.62 Rank: 6/144 |
4.682 Rank: 6/144 |
0.355 Rank: 1/144 |
0.78550 (±0.00157) Rank: 1/144 |
0.69405 Rank: 2/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 8k features, first using the oanet trained from scratch then guided-matching, setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00015 Root-SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
7860.73 | 274.85 Rank: 122/144 |
0.472 Rank: 97/144 |
0.845 Rank: 57/144 |
0.48887 (±0.00011) Rank: 89/144 |
437.74 Rank: 82/144 |
3814.80 Rank: 84/144 |
4.151 Rank: 115/144 |
0.458 Rank: 98/144 |
0.65067 (±0.00155) Rank: 104/144 |
0.56977 Rank: 98/144 |
Challenge organizers (contact) | sift-lowth | rootsift (128 float32: 512 bytes) | Root-SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00102 KeyNet-HardNet-8kSize: 512 bytes. Matches: built-in |
7997.63 | 815.40 Rank: 10/144 |
0.582 Rank: 2/144 |
0.788 Rank: 117/144 |
0.47395 (±0.00073) Rank: 97/144 |
356.21 Rank: 110/144 |
3366.01 Rank: 106/144 |
4.319 Rank: 81/144 |
0.464 Rank: 106/144 |
0.64829 (±0.00174) Rank: 105/144 |
0.56112 Rank: 104/144 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00519 SOSNet-Upright-AdaLAMSize: 512 bytes. Matches: custom |
7830.09 | 803.50 Rank: 11/144 |
0.486 Rank: 33/144 |
0.818 Rank: 108/144 |
0.56056 (±0.00000) Rank: 25/144 |
827.85 Rank: 17/144 |
6157.62 Rank: 10/144 |
4.700 Rank: 5/144 |
0.381 Rank: 12/144 |
0.75917 (±0.00446) Rank: 9/144 |
0.65987 Rank: 18/144 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | Using upright SOSNet descriptors with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. Matches post-processed with DEGENSAC. | https://arxiv.org/abs/2006.04250 | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00024 SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
7860.73 | 238.66 Rank: 129/144 |
0.472 Rank: 97/144 |
0.824 Rank: 99/144 |
0.45426 (±0.00097) Rank: 112/144 |
418.86 Rank: 88/144 |
3515.63 Rank: 97/144 |
4.001 Rank: 133/144 |
0.502 Rank: 125/144 |
0.60193 (±0.00185) Rank: 126/144 |
0.52810 Rank: 116/144 |
Challenge organizers (contact) | sift-lowth | sift (128 float32: 512 bytes) | SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00056 HardNet-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 439.36 Rank: 69/144 |
0.486 Rank: 45/144 |
0.864 Rank: 28/144 |
0.54393 (±0.00027) Rank: 39/144 |
480.13 Rank: 72/144 |
4167.34 Rank: 63/144 |
4.566 Rank: 31/144 |
0.417 Rank: 44/144 |
0.71542 (±0.00274) Rank: 44/144 |
0.62967 Rank: 41/144 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00047 GeoDesc w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 281.36 Rank: 120/144 |
0.472 Rank: 102/144 |
0.838 Rank: 75/144 |
0.47259 (±0.00033) Rank: 98/144 |
350.43 Rank: 117/144 |
3601.67 Rank: 88/144 |
4.234 Rank: 99/144 |
0.447 Rank: 87/144 |
0.65553 (±0.00444) Rank: 102/144 |
0.56406 Rank: 101/144 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00072 L2-Net-Upright w/ DEGENSAC (no F...Size: 512 bytes. Matches: built-in |
7829.63 | 435.68 Rank: 70/144 |
0.486 Rank: 45/144 |
0.859 Rank: 31/144 |
0.54497 (±0.00003) Rank: 36/144 |
395.53 Rank: 98/144 |
3603.85 Rank: 86/144 |
4.382 Rank: 64/144 |
0.452 Rank: 94/144 |
0.68491 (±0.00338) Rank: 81/144 |
0.61494 Rank: 51/144 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00518 UprightRootSIFT-AdaLAMSize: 512 bytes. Matches: custom |
6449.42 | 435.53 Rank: 71/144 |
0.436 Rank: 131/144 |
0.825 Rank: 98/144 |
0.49625 (±0.00000) Rank: 82/144 |
449.49 Rank: 78/144 |
4353.56 Rank: 57/144 |
4.453 Rank: 56/144 |
0.397 Rank: 24/144 |
0.72490 (±0.00190) Rank: 33/144 |
0.61057 Rank: 59/144 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Using upright RootSIFT with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. Matches post-processed with DEGENSAC. | https://arxiv.org/abs/2006.04250 | N/A | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00526 ASLFeat-MSSize: 512 bytes. Matches: built-in |
7384.19 | 456.55 Rank: 64/144 |
0.578 Rank: 5/144 |
0.752 Rank: 126/144 |
0.42208 (±0.00054) Rank: 116/144 |
532.83 Rank: 55/144 |
4002.80 Rank: 74/144 |
4.521 Rank: 45/144 |
0.447 Rank: 86/144 |
0.66568 (±0.00347) Rank: 94/144 |
0.54388 Rank: 113/144 |
Zixin Luo, Jiahui Zhang (contact) | aslfeat-ms | aslfeat-ms (128 float32: 512 bytes) | ASLFeat (joint predition of detectors and descriptors with multi-scale input) with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-04-27 | is_submission, is_challenge_2020 | |
Submission ID: 00627 Multiple kernel local descriptorSize: 512 bytes. Matches: built-in |
7862.74 | 354.57 Rank: 98/144 |
0.472 Rank: 91/144 |
0.845 Rank: 60/144 |
0.50919 (±0.00037) Rank: 75/144 |
580.16 Rank: 41/144 |
4784.46 Rank: 29/144 |
4.210 Rank: 104/144 |
0.436 Rank: 73/144 |
0.68323 (±0.00354) Rank: 83/144 |
0.59621 Rank: 79/144 |
(contact) | sift | mkdlw (128 float32: 512 bytes) | based on the paper [Understanding and Improving Kernel Local Descriptors] | https://arxiv.org/pdf/1811.11147 | https://arxiv.org/abs/1811.11147 | 20-10-28 | is_submission | |
Submission ID: 00143 CV-DoG-MKD-Concat-magsacSize: 512 bytes. Matches: built-in |
7860.77 | 381.37 Rank: 87/144 |
0.472 Rank: 82/144 |
0.833 Rank: 82/144 |
0.48096 (±0.00051) Rank: 96/144 |
348.03 Rank: 119/144 |
3507.39 Rank: 98/144 |
4.169 Rank: 112/144 |
0.471 Rank: 112/144 |
0.64763 (±0.00344) Rank: 108/144 |
0.56429 Rank: 100/144 |
Challenge organizers (contact) | sift | mkd-concat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the MKD-Concat descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1811.11147 | N/A | 21-02-05 | is_baseline | |
Submission ID: 00083 LogPolar-Upright w/ DEGENSAC (no...Size: 512 bytes. Matches: built-in |
7829.63 | 543.18 Rank: 42/144 |
0.486 Rank: 45/144 |
0.865 Rank: 27/144 |
0.55102 (±0.00015) Rank: 33/144 |
505.37 Rank: 68/144 |
4414.11 Rank: 52/144 |
4.518 Rank: 46/144 |
0.422 Rank: 54/144 |
0.71092 (±0.00251) Rank: 46/144 |
0.63097 Rank: 39/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00053 HardNet w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 575.07 Rank: 39/144 |
0.472 Rank: 102/144 |
0.842 Rank: 65/144 |
0.55022 (±0.00001) Rank: 34/144 |
426.84 Rank: 84/144 |
4001.40 Rank: 75/144 |
4.339 Rank: 79/144 |
0.419 Rank: 50/144 |
0.70962 (±0.00191) Rank: 49/144 |
0.62992 Rank: 40/144 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00557 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 614.16 Rank: 24/144 |
0.368 Rank: 135/144 |
0.513 Rank: 137/144 |
0.05482 (±0.00000) Rank: 138/144 |
626.57 Rank: 31/144 |
1793.00 Rank: 138/144 |
4.314 Rank: 86/144 |
0.517 Rank: 129/144 |
0.56933 (±0.00268) Rank: 132/144 |
0.31207 Rank: 134/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 10) [Mishkin 2019] | N/A | N/A | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00113 VL-Hess-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 290.24 Rank: 119/144 |
0.547 Rank: 12/144 |
0.814 Rank: 111/144 |
0.44501 (±0.00077) Rank: 114/144 |
347.39 Rank: 122/144 |
3209.10 Rank: 114/144 |
4.126 Rank: 117/144 |
0.519 Rank: 131/144 |
0.58657 (±0.00120) Rank: 129/144 |
0.51579 Rank: 120/144 |
Challenge organizers (contact) | hessian | vlsift (128 float32: 512 bytes) | VL-Hess-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00130 CV-DoG-AffNet-HardNet-kornia-DEG...Size: 512 bytes. Matches: built-in |
7833.97 | 403.41 Rank: 79/144 |
0.486 Rank: 74/144 |
0.883 Rank: 2/144 |
0.54468 (±0.00090) Rank: 37/144 |
580.47 Rank: 38/144 |
4671.35 Rank: 37/144 |
4.565 Rank: 33/144 |
0.402 Rank: 26/144 |
0.72668 (±0.00115) Rank: 30/144 |
0.63568 Rank: 34/144 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00533 upright-sift8k-hardnetSize: 512 bytes. Matches: built-in |
6589.88 | 371.42 Rank: 90/144 |
0.467 Rank: 123/144 |
0.852 Rank: 43/144 |
0.52463 (±0.00070) Rank: 56/144 |
404.02 Rank: 92/144 |
3501.62 Rank: 101/144 |
4.483 Rank: 52/144 |
0.428 Rank: 63/144 |
0.70123 (±0.00081) Rank: 59/144 |
0.61293 Rank: 54/144 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, harnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00541 sift and sosnet64 train scale(12...Size: 512 bytes. Matches: built-in |
7830.09 | 603.56 Rank: 30/144 |
0.486 Rank: 33/144 |
0.866 Rank: 24/144 |
0.58268 (±0.00061) Rank: 13/144 |
957.95 Rank: 10/144 |
6312.38 Rank: 8/144 |
4.607 Rank: 21/144 |
0.385 Rank: 14/144 |
0.74476 (±0.00137) Rank: 22/144 |
0.66372 Rank: 15/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-sos (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 348000), FLANN disabled | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00563 HardNet64-train-all-SOS-812000Size: 512 bytes. Matches: built-in |
7830.09 | 612.51 Rank: 25/144 |
0.486 Rank: 33/144 |
0.869 Rank: 21/144 |
0.58664 (±0.00095) Rank: 9/144 |
972.47 Rank: 8/144 |
6369.03 Rank: 4/144 |
4.612 Rank: 19/144 |
0.388 Rank: 18/144 |
0.74668 (±0.00251) Rank: 18/144 |
0.66666 Rank: 13/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-sos-812000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 812000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00105 VL-DoG-SIFT-8kSize: 512 bytes. Matches: built-in |
7880.59 | 261.60 Rank: 125/144 |
0.490 Rank: 20/144 |
0.826 Rank: 97/144 |
0.46555 (±0.00055) Rank: 104/144 |
324.62 Rank: 129/144 |
3030.67 Rank: 121/144 |
4.173 Rank: 109/144 |
0.461 Rank: 101/144 |
0.62829 (±0.00112) Rank: 116/144 |
0.54692 Rank: 108/144 |
Challenge organizers (contact) | dog | vlsift (128 float32: 512 bytes) | VL-DoG-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00110 VL-DoGAff-SIFT-8kSize: 512 bytes. Matches: built-in |
7892.05 | 317.13 Rank: 114/144 |
0.482 Rank: 79/144 |
0.820 Rank: 105/144 |
0.46657 (±0.00079) Rank: 103/144 |
311.54 Rank: 134/144 |
3061.54 Rank: 118/144 |
4.105 Rank: 121/144 |
0.475 Rank: 115/144 |
0.62964 (±0.00428) Rank: 112/144 |
0.54810 Rank: 107/144 |
Challenge organizers (contact) | dogaffine | vlsift (128 float32: 512 bytes) | VL-DoGAff-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00598 HarrisZ (more kpt) + DeepOrienta...Size: 512 bytes. Matches: custom |
4478.94 | 563.62 Rank: 40/144 |
0.455 Rank: 127/144 |
0.711 Rank: 128/144 |
0.39767 (±0.00000) Rank: 123/144 |
576.83 Rank: 43/144 |
2707.54 Rank: 131/144 |
4.381 Rank: 66/144 |
0.445 Rank: 85/144 |
0.69090 (±0.00163) Rank: 75/144 |
0.54428 Rank: 112/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ (start scale 2) [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.5, degeneracy check true, confidence 0.98, max iter 500000) [Mishkin 2019] | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00082 LogPolar w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 496.36 Rank: 55/144 |
0.472 Rank: 102/144 |
0.817 Rank: 109/144 |
0.49129 (±0.00011) Rank: 84/144 |
399.70 Rank: 95/144 |
4036.37 Rank: 72/144 |
4.340 Rank: 77/144 |
0.429 Rank: 66/144 |
0.69100 (±0.00154) Rank: 73/144 |
0.59115 Rank: 83/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00610 Hardnet-Upright-AdaLAMSize: 512 bytes. Matches: custom |
6556.61 | 627.71 Rank: 19/144 |
0.442 Rank: 130/144 |
0.828 Rank: 93/144 |
0.58300 (±0.00000) Rank: 12/144 |
645.47 Rank: 28/144 |
5074.91 Rank: 22/144 |
4.575 Rank: 30/144 |
0.361 Rank: 3/144 |
0.77056 (±0.00064) Rank: 3/144 |
0.67678 Rank: 4/144 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | Using upright Hardnet descriptors with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. Matches post-processed with DEGENSAC. | https://arxiv.org/abs/2006.04250 | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00568 Guided-SOSNet-lib-pSize: 512 bytes. Matches: custom |
7829.63 | 508.45 Rank: 52/144 |
0.486 Rank: 45/144 |
0.874 Rank: 12/144 |
0.57982 (±0.00000) Rank: 16/144 |
524.66 Rank: 57/144 |
4618.86 Rank: 41/144 |
4.632 Rank: 16/144 |
0.369 Rank: 6/144 |
0.75888 (±0.00437) Rank: 10/144 |
0.66935 Rank: 9/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhiwei Ruan, Zhongkun Chen, Jingchao Zhou, Pengfei Xu (contact) | sift8k | sosnet-lib-p (128 float32: 512 bytes) | sift and sosnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-15 | is_submission, is_challenge_2020 | |
Submission ID: 00117 VL-HessAffNet-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 299.02 Rank: 116/144 |
0.577 Rank: 6/144 |
0.831 Rank: 84/144 |
0.46793 (±0.00035) Rank: 101/144 |
350.69 Rank: 114/144 |
3327.71 Rank: 109/144 |
4.076 Rank: 124/144 |
0.487 Rank: 118/144 |
0.60691 (±0.00338) Rank: 123/144 |
0.53742 Rank: 114/144 |
Challenge organizers (contact) | hessian | affnetvlsift (128 float32: 512 bytes) | VL-HessAffNet-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00139 CV-DoG-TFeat-kornia-PyRANSACSize: 512 bytes. Matches: built-in |
7860.77 | 160.77 Rank: 140/144 |
0.472 Rank: 82/144 |
0.839 Rank: 71/144 |
0.40079 (±0.00116) Rank: 119/144 |
265.53 Rank: 137/144 |
2905.25 Rank: 125/144 |
4.038 Rank: 128/144 |
0.488 Rank: 122/144 |
0.62608 (±0.00138) Rank: 119/144 |
0.51344 Rank: 123/144 |
Challenge organizers (contact) | sift | tfeat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the TFeat descriptor. Implementation: OpenCV + kornia library | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | 21-02-05 | is_baseline | |
Submission ID: 00567 Guided-HardNet32-v1-lib-qht-pSize: 512 bytes. Matches: custom |
7829.63 | 520.40 Rank: 48/144 |
0.486 Rank: 45/144 |
0.875 Rank: 10/144 |
0.58509 (±0.00000) Rank: 10/144 |
536.93 Rank: 54/144 |
4685.16 Rank: 36/144 |
4.645 Rank: 12/144 |
0.376 Rank: 9/144 |
0.75702 (±0.00159) Rank: 12/144 |
0.67105 Rank: 7/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhiwei Ruan, Zhongkun Chen, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet32-v1-lib-qht-p (128 float32: 512 bytes) | sift and sosnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-15 | is_submission, is_challenge_2020 | |
Submission ID: 00077 SOSNet-Upright w/ DEGENSAC (no F...Size: 512 bytes. Matches: built-in |
7829.63 | 508.38 Rank: 53/144 |
0.486 Rank: 45/144 |
0.877 Rank: 6/144 |
0.57385 (±0.00041) Rank: 18/144 |
600.23 Rank: 35/144 |
4765.43 Rank: 31/144 |
4.584 Rank: 27/144 |
0.403 Rank: 30/144 |
0.73027 (±0.00284) Rank: 25/144 |
0.65206 Rank: 24/144 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-23 | is_baseline | |
Submission ID: 00101 KeyNet-HardNet-8kSize: 512 bytes. Matches: built-in |
7997.63 | 598.28 Rank: 32/144 |
0.582 Rank: 2/144 |
0.826 Rank: 96/144 |
0.49856 (±0.00038) Rank: 81/144 |
356.21 Rank: 110/144 |
3366.01 Rank: 106/144 |
4.319 Rank: 81/144 |
0.464 Rank: 107/144 |
0.64829 (±0.00174) Rank: 105/144 |
0.57342 Rank: 95/144 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00556 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 385.16 Rank: 84/144 |
0.368 Rank: 135/144 |
0.646 Rank: 134/144 |
0.04192 (±0.00000) Rank: 141/144 |
392.52 Rank: 104/144 |
1475.19 Rank: 141/144 |
4.454 Rank: 55/144 |
0.601 Rank: 135/144 |
0.48498 (±0.00118) Rank: 135/144 |
0.26345 Rank: 139/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 5) [Mishkin 2019] | N/A | N/A | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00524 SIFT-patchSize: 512 bytes. Matches: built-in |
7861.62 | 84.08 Rank: 143/144 |
0.472 Rank: 99/144 |
0.574 Rank: 136/144 |
0.22805 (±0.00076) Rank: 133/144 |
81.93 Rank: 143/144 |
991.42 Rank: 143/144 |
1.934 Rank: 144/144 |
0.687 Rank: 140/144 |
0.08662 (±0.00384) Rank: 143/144 |
0.15733 Rank: 142/144 |
feyman_priv (contact) | sift8k | l2net-arcface (128 float32: 512 bytes) | sift8k with l2net trained on google-landmark-dataset-v1(in 1000 class) | N/A | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00546 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 1020.86 Rank: 6/144 |
0.368 Rank: 135/144 |
0.366 Rank: 140/144 |
0.07780 (±0.00000) Rank: 136/144 |
1047.01 Rank: 7/144 |
2197.09 Rank: 135/144 |
4.023 Rank: 131/144 |
0.509 Rank: 127/144 |
0.59222 (±0.00276) Rank: 127/144 |
0.33501 Rank: 131/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 15, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-02 | is_baseline | |
Submission ID: 00635 MKDNet-indepSize: 512 bytes. Matches: built-in |
7862.74 | 336.51 Rank: 107/144 |
0.472 Rank: 91/144 |
0.848 Rank: 50/144 |
0.51519 (±0.00069) Rank: 67/144 |
550.86 Rank: 48/144 |
4772.22 Rank: 30/144 |
4.265 Rank: 93/144 |
0.427 Rank: 62/144 |
0.70004 (±0.00374) Rank: 62/144 |
0.60762 Rank: 64/144 |
(contact) | sift | mkdnet-indep (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-13 | is_submission | |
Submission ID: 00709 DISK (LCC/depth)Size: 512 bytes. Matches: built-in |
7844.17 | 1238.52 Rank: 2/144 |
0.644 Rank: 1/144 |
0.852 Rank: 41/144 |
0.55847 (±0.00084) Rank: 27/144 |
1663.81 Rank: 1/144 |
7483.98 Rank: 2/144 |
5.922 Rank: 1/144 |
0.391 Rank: 23/144 |
0.75024 (±0.00316) Rank: 15/144 |
0.65435 Rank: 21/144 |
Under review! NeurIPS anonymous submission, ID 1194 (To be released: 21-06-02) | disk-cc-continued-20-imsize-1024-nms-3-nump-8000 | disk-cc-continued-20-imsize-1024-nms-3-nump-8000 (128 float32: 512 bytes) | Local feature model learned via policy gradient. Model trained with a cycle-consistency loss and supervised with depth. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 3x3 window. We take the top 8000 features by score. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020, is_under_review | |
Submission ID: 00535 hardnetlib32P-Upright w/ DEGENSA...Size: 512 bytes. Matches: built-in |
7829.63 | 527.64 Rank: 46/144 |
0.486 Rank: 45/144 |
0.876 Rank: 8/144 |
0.57327 (±0.00078) Rank: 20/144 |
506.73 Rank: 67/144 |
4230.98 Rank: 60/144 |
4.517 Rank: 48/144 |
0.418 Rank: 46/144 |
0.72171 (±0.00156) Rank: 39/144 |
0.64749 Rank: 28/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Jingchao Zhou, Pengfei Xu (contact) | sift8k-no-dups-mps | hardnetlib32P (128 float32: 512 bytes) | sift and hardnet with 8k features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-27 | is_submission, is_challenge_2020 | |
Submission ID: 00108 VL-DoGAff-SIFT-8kSize: 512 bytes. Matches: built-in |
7892.05 | 171.63 Rank: 137/144 |
0.482 Rank: 79/144 |
0.848 Rank: 48/144 |
0.39839 (±0.00122) Rank: 122/144 |
311.54 Rank: 134/144 |
3061.54 Rank: 118/144 |
4.105 Rank: 121/144 |
0.475 Rank: 116/144 |
0.62964 (±0.00428) Rank: 112/144 |
0.51401 Rank: 122/144 |
Challenge organizers (contact) | dogaffine | vlsift (128 float32: 512 bytes) | VL-DoGAff-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00104 VL-DoG-SIFT-8kSize: 512 bytes. Matches: built-in |
7880.59 | 179.68 Rank: 136/144 |
0.490 Rank: 20/144 |
0.835 Rank: 80/144 |
0.39987 (±0.00068) Rank: 120/144 |
324.62 Rank: 129/144 |
3030.67 Rank: 121/144 |
4.173 Rank: 109/144 |
0.467 Rank: 111/144 |
0.62829 (±0.00112) Rank: 116/144 |
0.51408 Rank: 121/144 |
Challenge organizers (contact) | dog | vlsift (128 float32: 512 bytes) | VL-DoG-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00064 SOSNet w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 444.55 Rank: 67/144 |
0.472 Rank: 102/144 |
0.831 Rank: 85/144 |
0.51402 (±0.00057) Rank: 69/144 |
440.21 Rank: 80/144 |
4178.82 Rank: 61/144 |
4.378 Rank: 69/144 |
0.422 Rank: 55/144 |
0.70078 (±0.00375) Rank: 60/144 |
0.60740 Rank: 66/144 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00590 Guided-HardNet-epoch4Size: 512 bytes. Matches: custom |
7829.63 | 586.24 Rank: 34/144 |
0.486 Rank: 45/144 |
0.875 Rank: 9/144 |
0.59919 (±0.00000) Rank: 3/144 |
604.81 Rank: 34/144 |
5062.27 Rank: 23/144 |
4.710 Rank: 3/144 |
0.369 Rank: 7/144 |
0.76219 (±0.00253) Rank: 7/144 |
0.68069 Rank: 3/144 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-epoch4 (128 float32: 512 bytes) | sift and hardnet with 8k features, using the modified guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-24 | is_submission, is_challenge_2020 | |
Submission ID: 00506 sift and hardnet64Size: 512 bytes. Matches: built-in |
7861.62 | 358.55 Rank: 95/144 |
0.472 Rank: 100/144 |
0.849 Rank: 47/144 |
0.52085 (±0.00009) Rank: 60/144 |
585.33 Rank: 37/144 |
4957.86 Rank: 25/144 |
4.294 Rank: 87/144 |
0.419 Rank: 49/144 |
0.70472 (±0.00257) Rank: 56/144 |
0.61279 Rank: 55/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64 (128 float32: 512 bytes) | SIFT with 8000 keypoints, hardnet64 with 128 descriptors | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00116 VL-HessAffNet-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 209.26 Rank: 133/144 |
0.577 Rank: 6/144 |
0.841 Rank: 69/144 |
0.39330 (±0.00052) Rank: 124/144 |
350.69 Rank: 114/144 |
3327.71 Rank: 109/144 |
4.076 Rank: 124/144 |
0.490 Rank: 124/144 |
0.60691 (±0.00338) Rank: 123/144 |
0.50011 Rank: 126/144 |
Challenge organizers (contact) | hessian | affnetvlsift (128 float32: 512 bytes) | VL-HessAffNet-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00085 LogPolar-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 611.19 Rank: 26/144 |
0.486 Rank: 45/144 |
0.830 Rank: 88/144 |
0.51287 (±0.00038) Rank: 71/144 |
483.93 Rank: 70/144 |
4405.57 Rank: 55/144 |
4.542 Rank: 42/144 |
0.423 Rank: 57/144 |
0.70680 (±0.00247) Rank: 53/144 |
0.60983 Rank: 62/144 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-23 | is_baseline | |
Submission ID: 00109 VL-DoGAff-SIFT-8kSize: 512 bytes. Matches: built-in |
7892.05 | 250.11 Rank: 127/144 |
0.482 Rank: 79/144 |
0.838 Rank: 74/144 |
0.46795 (±0.00075) Rank: 100/144 |
311.54 Rank: 134/144 |
3061.54 Rank: 118/144 |
4.105 Rank: 121/144 |
0.475 Rank: 114/144 |
0.62964 (±0.00428) Rank: 112/144 |
0.54880 Rank: 106/144 |
Challenge organizers (contact) | dogaffine | vlsift (128 float32: 512 bytes) | VL-DoGAff-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00060 L2-Net w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 386.81 Rank: 83/144 |
0.472 Rank: 102/144 |
0.813 Rank: 112/144 |
0.49054 (±0.00043) Rank: 87/144 |
314.54 Rank: 132/144 |
3314.19 Rank: 112/144 |
4.204 Rank: 107/144 |
0.465 Rank: 108/144 |
0.65665 (±0.00287) Rank: 99/144 |
0.57360 Rank: 94/144 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00542 sift and hardnet64 train scale(1...Size: 512 bytes. Matches: built-in |
7830.09 | 607.44 Rank: 28/144 |
0.486 Rank: 33/144 |
0.872 Rank: 14/144 |
0.58801 (±0.00037) Rank: 6/144 |
950.74 Rank: 11/144 |
6286.36 Rank: 9/144 |
4.626 Rank: 17/144 |
0.390 Rank: 21/144 |
0.74662 (±0.00087) Rank: 19/144 |
0.66732 Rank: 12/144 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-l2 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 138000), FLANN disabled | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00569 Upright-SIFT-HardNetSize: 512 bytes. Matches: built-in |
6667.86 | 376.15 Rank: 89/144 |
0.491 Rank: 18/144 |
0.844 Rank: 62/144 |
0.51560 (±0.00043) Rank: 66/144 |
397.09 Rank: 97/144 |
3417.56 Rank: 105/144 |
4.479 Rank: 53/144 |
0.448 Rank: 92/144 |
0.68803 (±0.00532) Rank: 80/144 |
0.60181 Rank: 71/144 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, harnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-05-20 | is_submission, is_challenge_2020 | |
Submission ID: 00118 VL-HessAffNet-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 350.03 Rank: 99/144 |
0.577 Rank: 6/144 |
0.824 Rank: 100/144 |
0.46263 (±0.00079) Rank: 108/144 |
350.69 Rank: 114/144 |
3327.71 Rank: 109/144 |
4.076 Rank: 124/144 |
0.489 Rank: 123/144 |
0.60691 (±0.00338) Rank: 123/144 |
0.53477 Rank: 115/144 |
Challenge organizers (contact) | hessian | affnetvlsift (128 float32: 512 bytes) | VL-HessAffNet-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00058 HardNet-Upright w/ MAGSAC (no FL...Size: 512 bytes. Matches: built-in |
7829.63 | 707.86 Rank: 16/144 |
0.486 Rank: 45/144 |
0.853 Rank: 36/144 |
0.56374 (±0.00069) Rank: 24/144 |
509.07 Rank: 63/144 |
4250.40 Rank: 58/144 |
4.548 Rank: 39/144 |
0.412 Rank: 39/144 |
0.72309 (±0.00141) Rank: 35/144 |
0.64341 Rank: 29/144 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00131 CV-DoG-AffNet-HardNet-kornia-PyR...Size: 512 bytes. Matches: built-in |
7833.97 | 267.89 Rank: 123/144 |
0.486 Rank: 74/144 |
0.890 Rank: 1/144 |
0.45050 (±0.00103) Rank: 113/144 |
580.47 Rank: 38/144 |
4671.35 Rank: 37/144 |
4.565 Rank: 33/144 |
0.403 Rank: 29/144 |
0.72668 (±0.00115) Rank: 30/144 |
0.58859 Rank: 84/144 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00549 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 216.23 Rank: 132/144 |
0.284 Rank: 140/144 |
0.456 Rank: 138/144 |
0.14544 (±0.00000) Rank: 135/144 |
221.40 Rank: 141/144 |
1839.87 Rank: 137/144 |
3.608 Rank: 136/144 |
0.641 Rank: 137/144 |
0.47073 (±0.00158) Rank: 137/144 |
0.30808 Rank: 135/144 |
Fabio Bellavia (contact) | sift | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.75, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-03 | is_submission, is_challenge_2020 | |
Submission ID: 00607 HarrisZ (more kpt) + DeepOrienta...Size: 512 bytes. Matches: custom |
4478.94 | 746.02 Rank: 14/144 |
0.455 Rank: 127/144 |
0.699 Rank: 129/144 |
0.40725 (±0.00000) Rank: 117/144 |
763.44 Rank: 24/144 |
2953.35 Rank: 124/144 |
4.414 Rank: 57/144 |
0.442 Rank: 77/144 |
0.69399 (±0.00179) Rank: 69/144 |
0.55062 Rank: 105/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ (start scale 2) [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.75, degeneracy check true, confidence 0.98, max iter 500000) [Mishkin 2019] | N/A | N/A | 20-06-03 | is_submission, is_challenge_2020 | |
Submission ID: 00112 VL-Hess-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 204.37 Rank: 135/144 |
0.547 Rank: 12/144 |
0.823 Rank: 101/144 |
0.36954 (±0.00081) Rank: 126/144 |
347.39 Rank: 122/144 |
3209.10 Rank: 114/144 |
4.126 Rank: 117/144 |
0.517 Rank: 130/144 |
0.58657 (±0.00120) Rank: 129/144 |
0.47806 Rank: 128/144 |
Challenge organizers (contact) | hessian | vlsift (128 float32: 512 bytes) | VL-Hess-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00545 structured feature descriptionSize: 512 bytes. Matches: built-in |
7830.21 | 411.68 Rank: 75/144 |
0.486 Rank: 43/144 |
0.865 Rank: 26/144 |
0.54442 (±0.00084) Rank: 38/144 |
524.05 Rank: 58/144 |
4467.23 Rank: 51/144 |
4.551 Rank: 38/144 |
0.412 Rank: 38/144 |
0.72729 (±0.00300) Rank: 29/144 |
0.63586 Rank: 33/144 |
Mahdi Abolfazli Esfahani, Han Wang (contact) | sift | sfd (128 float32: 512 bytes) | structured descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC | N/A | N/A | 20-04-29 | is_submission, is_challenge_2020 | |
Submission ID: 00547 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 1078.85 Rank: 4/144 |
0.284 Rank: 140/144 |
0.166 Rank: 143/144 |
0.04936 (±0.00000) Rank: 140/144 |
1106.68 Rank: 5/144 |
2494.00 Rank: 133/144 |
3.400 Rank: 139/144 |
0.646 Rank: 138/144 |
0.43054 (±0.00185) Rank: 138/144 |
0.23995 Rank: 140/144 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 15, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-03 | is_submission, is_challenge_2020 | |
Submission ID: 00100 KeyNet-HardNet-8kSize: 512 bytes. Matches: built-in |
7997.63 | 448.11 Rank: 66/144 |
0.582 Rank: 2/144 |
0.838 Rank: 73/144 |
0.39971 (±0.00147) Rank: 121/144 |
356.21 Rank: 110/144 |
3366.01 Rank: 106/144 |
4.319 Rank: 81/144 |
0.462 Rank: 102/144 |
0.64829 (±0.00174) Rank: 105/144 |
0.52400 Rank: 118/144 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline |
Phototourism: restricted keypoints, standard descriptors (512 bytes)
Note: entries with the same multi-view configuration may seem duplicated. This is normal: performance is averaged across tasks.
Stereo | Multiview | Avg. | ||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Method | NF | NI | Rep. (3 pix.) |
MS (3 pix.) |
mAA (at 10o) |
NM | NL | TL | ATE | mAA (at 100) |
mAA (at 100) |
|||||||||
Submission ID: 00708 DISK (LCC/depth)Size: 512 bytes. Matches: built-in |
2048.00 | 404.19 Rank: 7/94 |
0.448 Rank: 11/94 |
0.852 Rank: 1/94 |
0.51315 (±0.00028) Rank: 8/94 |
527.48 Rank: 2/94 |
2428.04 Rank: 3/94 |
5.545 Rank: 5/94 |
0.410 Rank: 11/94 |
0.72705 (±0.00094) Rank: 11/94 |
0.62010 Rank: 9/94 |
Michal Tyszkiewicz (contact) | disk-cc-continued-20-imsize-1024-nms-3-nump-2048 | disk-cc-continued-20-imsize-1024-nms-3-nump-2048 (128 float32: 512 bytes) | Local feature model learned via policy gradient. Model trained with a cycle-consistency loss and supervised with depth. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 3x3 window. We take the top 2048 features by score. | N/A | N/A | 20-06-03 | is_submission, is_challenge_2020 | |
Submission ID: 00608 SIFT2k_2048_HardNet64-train-all-...Size: 512 bytes. Matches: custom |
2047.00 | 245.40 Rank: 22/94 |
0.346 Rank: 37/94 |
0.823 Rank: 6/94 |
0.49219 (±0.00000) Rank: 10/94 |
253.37 Rank: 40/94 |
1984.40 Rank: 16/94 |
4.607 Rank: 17/94 |
0.421 Rank: 15/94 |
0.70020 (±0.00213) Rank: 15/94 |
0.59619 Rank: 10/94 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | hardnet64-train-all-l2-138000-matched (128 float32: 512 bytes) | SIFT with 2048 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 138000), FLANN disabled, custom matches | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00702 L2-Net (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 117.11 Rank: 63/94 |
0.333 Rank: 51/94 |
0.808 Rank: 21/94 |
0.41918 (±0.00059) Rank: 42/94 |
179.79 Rank: 61/94 |
1361.34 Rank: 55/94 |
4.232 Rank: 57/94 |
0.481 Rank: 47/94 |
0.59682 (±0.00079) Rank: 56/94 |
0.50800 Rank: 47/94 |
Challenge organizers (contact) | sift-def | l2net-upright (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-06-01 | is_baseline | |
Submission ID: 00663 disk_degree(patch)_End-to-EndSize: 512 bytes. Matches: custom |
2048.00 | 448.22 Rank: 3/94 |
0.447 Rank: 15/94 |
0.834 Rank: 4/94 |
0.53645 (±0.00000) Rank: 6/94 |
458.32 Rank: 5/94 |
2377.71 Rank: 4/94 |
5.657 Rank: 2/94 |
0.373 Rank: 5/94 |
0.75520 (±0.00336) Rank: 5/94 |
0.64582 Rank: 6/94 |
Weiyue Zhao (contact) | disk | disk (128 float32: 512 bytes) | disk discriptors, followed by degree(patch)_End-to-End and DEGENSAC. | N/A | N/A | 21-04-28 | is_submission | |
Submission ID: 00653 sp_ae_sg_degensac_thSize: 512 bytes. Matches: custom |
2048.00 | 293.72 Rank: 14/94 |
0.352 Rank: 33/94 |
0.772 Rank: 56/94 |
0.54380 (±0.00000) Rank: 5/94 |
300.66 Rank: 24/94 |
1801.43 Rank: 23/94 |
4.723 Rank: 13/94 |
0.366 Rank: 4/94 |
0.75988 (±0.00277) Rank: 4/94 |
0.65184 Rank: 5/94 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-04-14 | is_submission | |
Submission ID: 00503 Guided matching hardnetSize: 512 bytes. Matches: custom |
1892.71 | 149.75 Rank: 47/94 |
0.333 Rank: 65/94 |
0.795 Rank: 34/94 |
0.42159 (±0.00000) Rank: 40/94 |
254.13 Rank: 38/94 |
1754.99 Rank: 27/94 |
4.283 Rank: 53/94 |
0.464 Rank: 35/94 |
0.64144 (±0.00207) Rank: 35/94 |
0.53152 Rank: 38/94 |
(contact) | sift | upright-hardnet (128 float32: 512 bytes) | Sift keypoint ; upright hardnet descriptors ; custom matching: use a deep learning based coarse matcher as a first step. In a second step match the keypoint according to descriptor distances but only for the matches that are close to the coarse match prediction | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00703 Upright Root-SIFT (OpenCV), DEGE...Size: 512 bytes. Matches: built-in |
1892.72 | 112.27 Rank: 71/94 |
0.333 Rank: 47/94 |
0.782 Rank: 44/94 |
0.39860 (±0.00077) Rank: 51/94 |
199.34 Rank: 52/94 |
1341.66 Rank: 57/94 |
4.090 Rank: 69/94 |
0.518 Rank: 64/94 |
0.56230 (±0.00234) Rank: 67/94 |
0.48045 Rank: 57/94 |
Challenge organizers (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Upright Root-SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-06-01 | is_baseline | |
Submission ID: 00644 Example: Upright SIFT (OpenCV)Size: 512 bytes. Matches: built-in |
1892.71 | 125.97 Rank: 61/94 |
0.333 Rank: 65/94 |
0.820 Rank: 11/94 |
0.43853 (±0.00030) Rank: 29/94 |
193.06 Rank: 57/94 |
1436.91 Rank: 46/94 |
4.278 Rank: 55/94 |
0.463 Rank: 33/94 |
0.63822 (±0.00469) Rank: 36/94 |
0.53838 Rank: 32/94 |
(contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | SIFT with 2048 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | https://opencv.org | 21-03-13 | is_submission | |
Submission ID: 00544 Key.Net + X-Net(Lib) w/ DEGENSACSize: 512 bytes. Matches: built-in |
2039.11 | 194.10 Rank: 28/94 |
0.447 Rank: 13/94 |
0.814 Rank: 15/94 |
0.42332 (±0.00096) Rank: 38/94 |
312.61 Rank: 20/94 |
1515.67 Rank: 39/94 |
4.567 Rank: 21/94 |
0.448 Rank: 24/94 |
0.66559 (±0.00233) Rank: 26/94 |
0.54445 Rank: 27/94 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet | x-net-lib (128 float32: 512 bytes) | N/A | N/A | 20-04-28 | is_submission, is_challenge_2020 | ||
Submission ID: 00603 SuperPoint-128d-masked + SuperGl...Size: 512 bytes. Matches: custom |
2041.09 | 404.73 Rank: 6/94 |
0.387 Rank: 25/94 |
0.774 Rank: 53/94 |
0.56769 (±0.00034) Rank: 3/94 |
415.75 Rank: 14/94 |
2158.03 Rank: 13/94 |
4.932 Rank: 10/94 |
0.357 Rank: 1/94 |
0.76987 (±0.00122) Rank: 3/94 |
0.66878 Rank: 3/94 |
Paul-Edouard Sarlin (contact) | superpoint-k2048-nms3-refine2-r1600forcecubic-masked-d.001 | superpoint-down128 (128 float32: 512 bytes) | SuperPoint detector (2048 keypoints, NMS with radius 3, confidence threshold 0.001, refinement, on 1600-pixel images). Detections on semantic classes sky and people are removed (segmentation from HFNetV2 trained on MIT ADE20K). SuperPoint descriptor, reduced to 128d with a linear autoencoder. SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.2 pixel inlier threshold). | https://arxiv.org/abs/1911.11763 | https://psarlin.com/superglue | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00622 Sift-HardNet-NM-Net_End-to-EndSize: 512 bytes. Matches: custom |
1892.70 | 152.13 Rank: 44/94 |
0.333 Rank: 67/94 |
0.807 Rank: 24/94 |
0.46620 (±0.00000) Rank: 14/94 |
156.54 Rank: 75/94 |
1403.97 Rank: 49/94 |
4.425 Rank: 32/94 |
0.443 Rank: 22/94 |
0.67160 (±0.00184) Rank: 24/94 |
0.56890 Rank: 18/94 |
Chen Zhao (contact) | siftdef | hardnet (128 float32: 512 bytes) | SIFT and HardNet, followed by NM-Net_End-to-End and DEGENSAC. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00668 sp_ae_sg_degensac_plusSize: 512 bytes. Matches: custom |
2048.00 | 394.88 Rank: 8/94 |
0.430 Rank: 19/94 |
0.749 Rank: 70/94 |
0.37510 (±0.00001) Rank: 64/94 |
402.56 Rank: 15/94 |
2246.23 Rank: 7/94 |
5.344 Rank: 7/94 |
0.432 Rank: 19/94 |
0.69832 (±0.00287) Rank: 17/94 |
0.53671 Rank: 35/94 |
(contact) | sid-00668-superpoint | sid-00668-superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-05-03 | is_submission | |
Submission ID: 00667 sp_ae_sg_degensac_xpSize: 512 bytes. Matches: custom |
2048.00 | 446.24 Rank: 4/94 |
0.407 Rank: 22/94 |
0.791 Rank: 37/94 |
0.58918 (±0.00001) Rank: 2/94 |
458.12 Rank: 6/94 |
2290.85 Rank: 6/94 |
5.123 Rank: 8/94 |
0.365 Rank: 3/94 |
0.77685 (±0.00244) Rank: 1/94 |
0.68301 Rank: 1/94 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-04-29 | is_submission | |
Submission ID: 00023 LogPolarDesc, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 116.83 Rank: 64/94 |
0.323 Rank: 75/94 |
0.770 Rank: 57/94 |
0.39335 (±0.00060) Rank: 53/94 |
162.84 Rank: 71/94 |
1385.60 Rank: 53/94 |
4.046 Rank: 72/94 |
0.519 Rank: 65/94 |
0.57149 (±0.00344) Rank: 63/94 |
0.48242 Rank: 55/94 |
Challenge organizers (contact) | sift-def | logpolar (128 float32: 512 bytes) | LogPolarDesc descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00583 R2D2Size: 512 bytes. Matches: built-in |
2048.00 | 169.38 Rank: 34/94 |
0.398 Rank: 24/94 |
0.695 Rank: 77/94 |
0.33715 (±0.00149) Rank: 74/94 |
306.15 Rank: 21/94 |
1235.83 Rank: 73/94 |
4.156 Rank: 67/94 |
0.496 Rank: 51/94 |
0.60223 (±0.00103) Rank: 53/94 |
0.46969 Rank: 68/94 |
(contact) | r2d2-5k-p-aug | r2d2-5k-p-aug (128 float32: 512 bytes) | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | ||
Submission ID: 00001 HardNet, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 109.52 Rank: 74/94 |
0.323 Rank: 75/94 |
0.776 Rank: 50/94 |
0.38578 (±0.00131) Rank: 59/94 |
153.29 Rank: 77/94 |
1306.64 Rank: 62/94 |
4.026 Rank: 76/94 |
0.539 Rank: 77/94 |
0.55573 (±0.00175) Rank: 73/94 |
0.47075 Rank: 66/94 |
Challenge organizers (contact) | sift-def | hardnet (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00585 NM-Net_v2Size: 512 bytes. Matches: custom |
1892.70 | 151.46 Rank: 46/94 |
0.333 Rank: 67/94 |
0.809 Rank: 19/94 |
0.46620 (±0.00000) Rank: 13/94 |
155.91 Rank: 76/94 |
1392.47 Rank: 52/94 |
4.424 Rank: 33/94 |
0.459 Rank: 30/94 |
0.66837 (±0.00125) Rank: 25/94 |
0.56729 Rank: 19/94 |
Chen Zhao (contact) | siftdef | hardnet (128 float32: 512 bytes) | SIFT and HardNet, followed by NM-Net_v2 and DEGENSAC. | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | |
Submission ID: 00589 SuperPoint-128d + SuperGlue + DE...Size: 512 bytes. Matches: custom |
1973.62 | 320.46 Rank: 13/94 |
0.364 Rank: 30/94 |
0.772 Rank: 55/94 |
0.55214 (±0.00098) Rank: 4/94 |
429.49 Rank: 13/94 |
2130.73 Rank: 14/94 |
4.570 Rank: 19/94 |
0.384 Rank: 7/94 |
0.75283 (±0.00086) Rank: 7/94 |
0.65248 Rank: 4/94 |
Paul-Edouard Sarlin (contact) | superpoint-k2048-nms3-refine2-r1600forcecubic | superpoint-down128 (128 float32: 512 bytes) | SuperPoint detector (2048 keypoints, NMS with radius 3, refinement, on 1600-pixel images) and descriptor; reduced to 128d with a linear autoencoder. SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.2 pixels inlier threshold). | https://arxiv.org/abs/1911.11763 | https://psarlin.com/superglue | 20-05-24 | is_submission, is_challenge_2020 | |
Submission ID: 00617 pffNet + SuperPoint + DEGENSACSize: 512 bytes. Matches: built-in |
1267.22 | 55.58 Rank: 93/94 |
0.341 Rank: 38/94 |
0.629 Rank: 84/94 |
0.25514 (±0.00059) Rank: 82/94 |
131.91 Rank: 88/94 |
893.88 Rank: 91/94 |
4.339 Rank: 37/94 |
0.534 Rank: 73/94 |
0.55935 (±0.00143) Rank: 71/94 |
0.40724 Rank: 80/94 |
Jongmin Lee, Seungwook Kim, Yoonwoo Jeong (contact) | superpoint | pffnet (128 float32: 512 bytes) | pffNet descriptors+ SuperPoint keypoints + DEGENSAC outlier-filtering | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00123 KeyNet-HardNet-2kSize: 512 bytes. Matches: built-in |
2048.00 | 134.39 Rank: 56/94 |
0.469 Rank: 1/94 |
0.818 Rank: 12/94 |
0.32725 (±0.00026) Rank: 77/94 |
195.33 Rank: 53/94 |
1276.30 Rank: 68/94 |
4.493 Rank: 28/94 |
0.490 Rank: 50/94 |
0.61606 (±0.00141) Rank: 47/94 |
0.47166 Rank: 65/94 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 2048 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-03 | is_baseline | |
Submission ID: 00008 L2-Net (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 138.44 Rank: 53/94 |
0.333 Rank: 51/94 |
0.786 Rank: 40/94 |
0.39980 (±0.00020) Rank: 50/94 |
171.93 Rank: 64/94 |
1333.56 Rank: 59/94 |
4.226 Rank: 59/94 |
0.496 Rank: 52/94 |
0.58762 (±0.00228) Rank: 57/94 |
0.49371 Rank: 54/94 |
Challenge organizers (contact) | sift-def | l2net-upright (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-22 | is_baseline | |
Submission ID: 00647 sp_ae_sg_ransacSize: 512 bytes. Matches: custom |
1873.96 | 252.53 Rank: 19/94 |
0.323 Rank: 72/94 |
0.663 Rank: 81/94 |
0.06318 (±0.00000) Rank: 93/94 |
256.66 Rank: 36/94 |
1663.09 Rank: 31/94 |
4.771 Rank: 11/94 |
0.563 Rank: 82/94 |
0.52605 (±0.00451) Rank: 78/94 |
0.29461 Rank: 86/94 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-03-29 | is_submission | |
Submission ID: 00646 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: built-in |
2048.00 | 241.83 Rank: 24/94 |
0.450 Rank: 4/94 |
0.689 Rank: 79/94 |
0.32109 (±0.00043) Rank: 79/94 |
301.64 Rank: 22/94 |
1593.08 Rank: 35/94 |
4.285 Rank: 51/94 |
0.497 Rank: 54/94 |
0.60774 (±0.00266) Rank: 49/94 |
0.46442 Rank: 71/94 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features score-wise. HardNet, pre-trained on all datasets, is used as a descriptor head. Graph-Cut(GC)-RANSAC is used as a robust estimator. Cyclic consistency matching with a threshold of 0.95 is used. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-03-28 | is_submission | |
Submission ID: 00509 Upright-Sift + X-Net-lib w/ DEGE...Size: 512 bytes. Matches: built-in |
1892.70 | 116.55 Rank: 67/94 |
0.333 Rank: 71/94 |
0.821 Rank: 9/94 |
0.42382 (±0.00089) Rank: 37/94 |
168.67 Rank: 67/94 |
1319.63 Rank: 61/94 |
4.303 Rank: 47/94 |
0.468 Rank: 40/94 |
0.62953 (±0.00216) Rank: 42/94 |
0.52668 Rank: 41/94 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | sift-def | x-net-lib-upright-no-dups (128 float32: 512 bytes) | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | ||
Submission ID: 00658 disk_degree_end-to-endSize: 512 bytes. Matches: custom |
2048.00 | 346.06 Rank: 11/94 |
0.447 Rank: 15/94 |
0.843 Rank: 2/94 |
0.51061 (±0.00000) Rank: 9/94 |
354.06 Rank: 17/94 |
2170.03 Rank: 10/94 |
5.679 Rank: 1/94 |
0.393 Rank: 8/94 |
0.74934 (±0.00266) Rank: 8/94 |
0.62998 Rank: 8/94 |
Weiyue Zhao (contact) | disk | disk (128 float32: 512 bytes) | disk discriptors, followed by degree_End-to-End and DEGENSAC. | N/A | N/A | 21-04-22 | is_submission | |
Submission ID: 00605 guided-hardnet-epoch2-v2Size: 512 bytes. Matches: custom |
2047.76 | 220.88 Rank: 25/94 |
0.347 Rank: 34/94 |
0.778 Rank: 48/94 |
0.46469 (±0.00000) Rank: 15/94 |
226.84 Rank: 44/94 |
1827.60 Rank: 20/94 |
4.512 Rank: 25/94 |
0.427 Rank: 17/94 |
0.69306 (±0.00160) Rank: 19/94 |
0.57888 Rank: 16/94 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift2k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 2k features, using the oanet and guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00645 sp_ae_sgSize: 512 bytes. Matches: custom |
1873.96 | 488.91 Rank: 2/94 |
0.323 Rank: 72/94 |
0.585 Rank: 86/94 |
0.28532 (±0.00000) Rank: 81/94 |
498.60 Rank: 3/94 |
2432.22 Rank: 2/94 |
4.542 Rank: 23/94 |
0.399 Rank: 10/94 |
0.72857 (±0.00361) Rank: 10/94 |
0.50694 Rank: 48/94 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-03-18 | is_submission | |
Submission ID: 00527 SEKDSize: 512 bytes. Matches: built-in |
1786.69 | 106.34 Rank: 76/94 |
0.383 Rank: 28/94 |
0.759 Rank: 66/94 |
0.38886 (±0.00013) Rank: 54/94 |
150.11 Rank: 79/94 |
1063.08 Rank: 88/94 |
4.575 Rank: 18/94 |
0.465 Rank: 37/94 |
0.63303 (±0.00244) Rank: 40/94 |
0.51095 Rank: 46/94 |
Yafei Song, Ling Cai, Mingyang Li (contact) | sekd | sekd (128 float32: 512 bytes) | The name of our method is Self-Evolving Keypoint Detection and Description (SEKD). Now, the SEKD model is trained only using COCO test images. In this submission each image has up to 2048 SEKD keypoints, and 128-dim float descriptor. We use the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-04-27 | is_submission, is_challenge_2020 | |
Submission ID: ????? CV-DoG-HardNet8-PTSize: 512 bytes. Matches: built-in |
2048.00 | 114.84 Rank: 70/94 |
0.321 Rank: 88/94 |
0.790 Rank: 38/94 |
0.38200 (±0.00081) Rank: 61/94 |
117.84 Rank: 94/94 |
1067.13 Rank: 87/94 |
4.051 Rank: 71/94 |
0.531 Rank: 71/94 |
0.53035 (±0.00531) Rank: 76/94 |
0.45617 Rank: 73/94 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift2k | h8e512pt (128 float32: 512 bytes) | [sid:00593] HardNet8 with PCA compression | N/A | N/A | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00044 Upright SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1892.72 | 104.22 Rank: 79/94 |
0.333 Rank: 47/94 |
0.763 Rank: 63/94 |
0.37078 (±0.00056) Rank: 66/94 |
205.17 Rank: 48/94 |
1300.49 Rank: 64/94 |
4.000 Rank: 81/94 |
0.563 Rank: 81/94 |
0.52549 (±0.00120) Rank: 79/94 |
0.44814 Rank: 75/94 |
Challenge organizers (contact) | sift-def | sift-upright (128 float32: 512 bytes) | Upright SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00004 SOSNet (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 176.54 Rank: 30/94 |
0.333 Rank: 51/94 |
0.782 Rank: 43/94 |
0.43803 (±0.00092) Rank: 31/94 |
183.92 Rank: 58/94 |
1403.66 Rank: 50/94 |
4.314 Rank: 43/94 |
0.474 Rank: 45/94 |
0.62181 (±0.00161) Rank: 45/94 |
0.52992 Rank: 40/94 |
Challenge organizers (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00043 DELF-GLD (128D), PyRANSACSize: 512 bytes. Matches: built-in |
2036.82 | 94.70 Rank: 82/94 |
0.109 Rank: 90/94 |
0.156 Rank: 90/94 |
0.07298 (±0.00056) Rank: 91/94 |
430.68 Rank: 11/94 |
2162.72 Rank: 11/94 |
2.535 Rank: 90/94 |
0.887 Rank: 91/94 |
0.13247 (±0.00116) Rank: 89/94 |
0.10272 Rank: 90/94 |
Challenge organizers (contact) | delf-gld-2k-128d | delf-gld-2k-128d (128 float32: 512 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 128 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with PyRANSAC (DEGENSAC with the degeneracy check turned off). | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-23 | is_baseline | |
Submission ID: 00612 SuperPoint-128d-adapt + SuperGlu...Size: 512 bytes. Matches: custom |
2048.00 | 441.49 Rank: 5/94 |
0.407 Rank: 21/94 |
0.789 Rank: 39/94 |
0.59034 (±0.00050) Rank: 1/94 |
452.99 Rank: 9/94 |
2245.43 Rank: 8/94 |
5.092 Rank: 9/94 |
0.358 Rank: 2/94 |
0.77337 (±0.00167) Rank: 2/94 |
0.68186 Rank: 2/94 |
Paul-Edouard Sarlin (contact) | superpoint-k2048-nms4-refine2-r1600forcecubic-masked-d.001-adapt50 | superpoint-down128 (128 float32: 512 bytes) | SuperPoint detector (2048 keypoints, NMS with radius 4, confidence threshold 0.001, refinement, on 1600-pixel images). The detection heatmap is improved with test-time homographic adaptation (50 iterations), and detections on semantic classes sky and people are removed (segmentation from HFNetV2 trained on MIT ADE20K). SuperPoint descriptor, reduced to 128d with a linear autoencoder. SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.1 pixel inlier threshold). | https://arxiv.org/abs/1911.11763 | https://psarlin.com/superglue | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00615 SIFT-Fusion_Max-NM-Net_End-to-En...Size: 512 bytes. Matches: custom |
1892.70 | 115.46 Rank: 68/94 |
0.333 Rank: 67/94 |
0.783 Rank: 42/94 |
0.40244 (±0.00000) Rank: 48/94 |
119.23 Rank: 93/94 |
1167.54 Rank: 80/94 |
4.367 Rank: 35/94 |
0.489 Rank: 49/94 |
0.60720 (±0.00114) Rank: 50/94 |
0.50482 Rank: 49/94 |
Chen Zhao (contact) | siftdef | fusion-max (128 float32: 512 bytes) | SIFT and Fusion_Max, followed by NM-Net_End-to-End and DEGENSAC. | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00002 SURF (OpenCV), DEGENSACSize: 256 bytes. Matches: built-in |
2010.88 | 71.90 Rank: 88/94 |
0.321 Rank: 87/94 |
0.634 Rank: 82/94 |
0.20761 (±0.00029) Rank: 85/94 |
300.02 Rank: 25/94 |
1208.06 Rank: 78/94 |
3.549 Rank: 87/94 |
0.695 Rank: 87/94 |
0.37845 (±0.00399) Rank: 87/94 |
0.29303 Rank: 87/94 |
Challenge organizers (contact) | surf-def | surf (64 float32: 256 bytes) | SURF with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00641 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: built-in |
2048.00 | 219.72 Rank: 26/94 |
0.450 Rank: 4/94 |
0.722 Rank: 75/94 |
0.34184 (±0.00032) Rank: 72/94 |
301.64 Rank: 22/94 |
1595.30 Rank: 34/94 |
4.278 Rank: 54/94 |
0.499 Rank: 55/94 |
0.60963 (±0.00152) Rank: 48/94 |
0.47574 Rank: 61/94 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features by score. HardNet ,pre-trained on all datasets, is used as a descriptor head. MAGSAC is used as a robust estimator. Cyclic consistency matching with a threshold of 0.95 is used. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-03-04 | is_submission | |
Submission ID: 00705 Upright SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1892.72 | 104.74 Rank: 78/94 |
0.333 Rank: 47/94 |
0.764 Rank: 62/94 |
0.37292 (±0.00081) Rank: 65/94 |
204.13 Rank: 49/94 |
1293.75 Rank: 65/94 |
4.008 Rank: 78/94 |
0.567 Rank: 83/94 |
0.52808 (±0.00416) Rank: 77/94 |
0.45050 Rank: 74/94 |
Challenge organizers (contact) | sift-def | sift-upright (128 float32: 512 bytes) | Upright SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-06-01 | is_baseline | |
Submission ID: 00014 L2-Net, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 107.38 Rank: 75/94 |
0.323 Rank: 81/94 |
0.766 Rank: 59/94 |
0.35649 (±0.00110) Rank: 70/94 |
138.97 Rank: 86/94 |
1211.49 Rank: 75/94 |
3.948 Rank: 82/94 |
0.538 Rank: 76/94 |
0.52019 (±0.00120) Rank: 80/94 |
0.43834 Rank: 77/94 |
Challenge organizers (contact) | sift-def | l2net (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-23 | is_baseline | |
Submission ID: 00032 Root-SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1936.30 | 83.46 Rank: 86/94 |
0.323 Rank: 85/94 |
0.756 Rank: 68/94 |
0.33605 (±0.00092) Rank: 75/94 |
167.84 Rank: 68/94 |
1141.88 Rank: 83/94 |
3.788 Rank: 85/94 |
0.576 Rank: 84/94 |
0.46826 (±0.00154) Rank: 84/94 |
0.40216 Rank: 81/94 |
Challenge organizers (contact) | sift-def | rootsift (128 float32: 512 bytes) | Root-SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00007 L2-Net, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 87.04 Rank: 85/94 |
0.323 Rank: 81/94 |
0.779 Rank: 47/94 |
0.36265 (±0.00090) Rank: 67/94 |
138.97 Rank: 86/94 |
1211.49 Rank: 75/94 |
3.948 Rank: 82/94 |
0.537 Rank: 74/94 |
0.52019 (±0.00120) Rank: 80/94 |
0.44142 Rank: 76/94 |
Challenge organizers (contact) | sift-def | l2net (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-23 | is_baseline | |
Submission ID: 00599 SEKDSize: 512 bytes. Matches: built-in |
2043.44 | 129.46 Rank: 59/94 |
0.386 Rank: 26/94 |
0.809 Rank: 20/94 |
0.45066 (±0.00057) Rank: 22/94 |
176.60 Rank: 62/94 |
1209.63 Rank: 77/94 |
4.437 Rank: 30/94 |
0.454 Rank: 27/94 |
0.66094 (±0.00227) Rank: 28/94 |
0.55580 Rank: 23/94 |
Yafei Song, Ling Cai, Jia Li, Yonghong Tian, Mingyang Li (contact) | sekd | sekd (128 float32: 512 bytes) | Our method named SEKD: Self-Evolving Keypoint Detection and Description, where the SEKD model is trained using COCO validation set. In this submission each image has up to 2048 SEKD keypoints, and 128-dim float descriptor. We use the built-in matcher (bidirectional filter with the both strategy, without flann, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-05-29 | is_submission, is_challenge_2020 | |
Submission ID: 00136 CV-DoG-AffNet-HardNet-kornia-MAG...Size: 512 bytes. Matches: built-in |
2047.84 | 195.17 Rank: 27/94 |
0.339 Rank: 40/94 |
0.774 Rank: 52/94 |
0.41747 (±0.00072) Rank: 43/94 |
284.06 Rank: 30/94 |
1788.70 Rank: 24/94 |
4.191 Rank: 63/94 |
0.511 Rank: 61/94 |
0.58536 (±0.00125) Rank: 59/94 |
0.50141 Rank: 51/94 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00011 GeoDesc, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 93.44 Rank: 83/94 |
0.323 Rank: 81/94 |
0.759 Rank: 65/94 |
0.34215 (±0.00054) Rank: 71/94 |
120.68 Rank: 91/94 |
1098.70 Rank: 85/94 |
4.002 Rank: 79/94 |
0.552 Rank: 80/94 |
0.50594 (±0.00073) Rank: 82/94 |
0.42405 Rank: 78/94 |
Challenge organizers (contact) | sift-def | geodesc (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-23 | is_baseline | |
Submission ID: 00018 R2D2 (r2d2-wasf-n16), DEGENSACSize: 512 bytes. Matches: built-in |
2048.00 | 176.49 Rank: 31/94 |
0.422 Rank: 20/94 |
0.725 Rank: 74/94 |
0.35783 (±0.00061) Rank: 69/94 |
262.02 Rank: 35/94 |
1188.11 Rank: 79/94 |
4.299 Rank: 50/94 |
0.499 Rank: 56/94 |
0.60639 (±0.00124) Rank: 51/94 |
0.48211 Rank: 56/94 |
Challenge organizers (contact) | r2d2-wasf-n16 | r2d2-wasf-n16 (128 float32: 512 bytes) | R2D2 (model r2d2-wasf-n16) with 2k features, using the built-in matcher (bidirectional filter with the both strategy) and no ratio test, and DEGENSAC with optimal parameters for stereo. | N/A | https://opencv.org | 20-04-23 | is_baseline | |
Submission ID: ????? CV-DoG-HardNet8-PTv2Size: 512 bytes. Matches: built-in |
2047.77 | 153.69 Rank: 42/94 |
0.339 Rank: 44/94 |
0.821 Rank: 8/94 |
0.43869 (±0.00017) Rank: 28/94 |
158.63 Rank: 74/94 |
1221.12 Rank: 74/94 |
4.350 Rank: 36/94 |
0.484 Rank: 48/94 |
0.60206 (±0.00003) Rank: 54/94 |
0.52038 Rank: 43/94 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift2k | h8e512pt (128 float32: 512 bytes) | [sid:00619] HardNet8 with PCA compression, batch sampling from few images | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00025 DELF-GLD (32D), DEGENSACSize: 128 bytes. Matches: built-in |
2036.82 | 96.56 Rank: 81/94 |
0.109 Rank: 90/94 |
0.119 Rank: 93/94 |
0.06573 (±0.00030) Rank: 92/94 |
453.19 Rank: 7/94 |
1805.48 Rank: 21/94 |
2.425 Rank: 92/94 |
0.920 Rank: 94/94 |
0.09163 (±0.00140) Rank: 92/94 |
0.07868 Rank: 92/94 |
Challenge organizers (contact) | delf-gld-2k-32d | delf-gld-2k-32d (32 float32: 128 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 32 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with DEGENSAC. | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00578 Key.Net-s + DescNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
2033.67 | 246.62 Rank: 21/94 |
0.449 Rank: 9/94 |
0.805 Rank: 25/94 |
0.45418 (±0.00056) Rank: 20/94 |
331.55 Rank: 19/94 |
1621.68 Rank: 33/94 |
4.570 Rank: 20/94 |
0.447 Rank: 23/94 |
0.67414 (±0.00371) Rank: 23/94 |
0.56416 Rank: 20/94 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet-s | descnet (128 float32: 512 bytes) | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | ||
Submission ID: 00016 LogPolarDesc, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 148.56 Rank: 48/94 |
0.323 Rank: 75/94 |
0.752 Rank: 69/94 |
0.38840 (±0.00064) Rank: 55/94 |
162.84 Rank: 71/94 |
1385.60 Rank: 53/94 |
4.046 Rank: 72/94 |
0.516 Rank: 63/94 |
0.57149 (±0.00344) Rank: 63/94 |
0.47994 Rank: 58/94 |
Challenge organizers (contact) | sift-def | logpolar (128 float32: 512 bytes) | LogPolarDesc descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-23 | is_baseline | |
Submission ID: 00704 SOSNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 171.22 Rank: 33/94 |
0.333 Rank: 51/94 |
0.804 Rank: 26/94 |
0.45053 (±0.00053) Rank: 23/94 |
194.04 Rank: 54/94 |
1442.32 Rank: 45/94 |
4.313 Rank: 45/94 |
0.467 Rank: 39/94 |
0.63586 (±0.00246) Rank: 38/94 |
0.54320 Rank: 28/94 |
Challenge organizers (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-06-02 | is_baseline | |
Submission ID: 00604 Key.Net+GIFT+GMCNet+DEGENSACSize: 512 bytes. Matches: custom |
2046.07 | 377.20 Rank: 9/94 |
0.450 Rank: 8/94 |
0.776 Rank: 49/94 |
0.45282 (±0.00000) Rank: 21/94 |
386.63 Rank: 16/94 |
1474.77 Rank: 42/94 |
4.663 Rank: 16/94 |
0.420 Rank: 13/94 |
0.70500 (±0.00294) Rank: 12/94 |
0.57891 Rank: 15/94 |
Yuan Liu (contact) | keynet-2k | scale-gift (128 float32: 512 bytes) | Detecting by Key.Net, descriptors from GIFT, matching by Graph Motion Coherence Network, geometry estimated by DEGENSAC with inlier threshold 0.7 | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00700 HardNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 152.69 Rank: 43/94 |
0.333 Rank: 51/94 |
0.812 Rank: 17/94 |
0.46093 (±0.00080) Rank: 17/94 |
201.26 Rank: 50/94 |
1467.86 Rank: 44/94 |
4.311 Rank: 46/94 |
0.466 Rank: 38/94 |
0.63544 (±0.00324) Rank: 39/94 |
0.54819 Rank: 25/94 |
Challenge organizers (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-06-01 | is_baseline | |
Submission ID: 00036 SOSNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 155.75 Rank: 40/94 |
0.333 Rank: 51/94 |
0.794 Rank: 35/94 |
0.43077 (±0.00017) Rank: 34/94 |
183.92 Rank: 58/94 |
1403.66 Rank: 50/94 |
4.314 Rank: 43/94 |
0.472 Rank: 42/94 |
0.62181 (±0.00161) Rank: 45/94 |
0.52629 Rank: 42/94 |
Challenge organizers (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00575 SIFT2k-2048-HardNet64-rain-all-s...Size: 512 bytes. Matches: built-in |
2047.77 | 161.77 Rank: 36/94 |
0.339 Rank: 43/94 |
0.814 Rank: 13/94 |
0.43543 (±0.00051) Rank: 32/94 |
255.45 Rank: 37/94 |
1708.29 Rank: 28/94 |
4.327 Rank: 41/94 |
0.451 Rank: 26/94 |
0.64653 (±0.00221) Rank: 31/94 |
0.54098 Rank: 29/94 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | sift2k-2048-hardnet64-train-all-sos-812000 (128 float32: 512 bytes) | SIFT with 2048 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 812000), FLANN disabled | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | |
Submission ID: 00531 Key.Net + X-Net w/ DEGENSACSize: 512 bytes. Matches: built-in |
2040.08 | 173.65 Rank: 32/94 |
0.447 Rank: 12/94 |
0.797 Rank: 33/94 |
0.38608 (±0.00068) Rank: 58/94 |
276.91 Rank: 34/94 |
1468.50 Rank: 43/94 |
4.447 Rank: 29/94 |
0.463 Rank: 34/94 |
0.64369 (±0.00081) Rank: 34/94 |
0.51489 Rank: 45/94 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet | x-net (128 float32: 512 bytes) | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | ||
Submission ID: 00565 SIFT2k_2000_HardNet64-train-all-...Size: 512 bytes. Matches: built-in |
1999.83 | 157.70 Rank: 38/94 |
0.336 Rank: 45/94 |
0.813 Rank: 16/94 |
0.43188 (±0.00005) Rank: 33/94 |
248.91 Rank: 42/94 |
1666.27 Rank: 30/94 |
4.323 Rank: 42/94 |
0.465 Rank: 36/94 |
0.64404 (±0.00271) Rank: 33/94 |
0.53796 Rank: 33/94 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | sift2k-2000-hardnet64-train-all-sos-812000 (128 float32: 512 bytes) | SIFT with 2000 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 812000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00512 Hardnet with DEGENSAC for stereoSize: 512 bytes. Matches: built-in |
1999.50 | 51.91 Rank: 94/94 |
0.172 Rank: 89/94 |
0.558 Rank: 88/94 |
0.20771 (±0.00049) Rank: 84/94 |
181.48 Rank: 60/94 |
1692.63 Rank: 29/94 |
3.050 Rank: 89/94 |
0.695 Rank: 86/94 |
0.38463 (±0.00275) Rank: 86/94 |
0.29617 Rank: 85/94 |
Vu Trung Nghia & Nguyen Trung Hieu (contact) | siftdef | hardnet-64 (128 float32: 512 bytes) | Using hardnet network to embed a patch (32 x 32) to a 128 (float32) dimensions vector, For stereo, we use DEGENSAC (Chum et al, CVPR'05) with optimal settings | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00012 GeoDesc (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 144.92 Rank: 49/94 |
0.333 Rank: 51/94 |
0.765 Rank: 61/94 |
0.38124 (±0.00056) Rank: 62/94 |
149.66 Rank: 80/94 |
1237.96 Rank: 71/94 |
4.219 Rank: 61/94 |
0.526 Rank: 70/94 |
0.56687 (±0.00261) Rank: 65/94 |
0.47405 Rank: 62/94 |
Challenge organizers (contact) | sift-def | geodesc-upright (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-23 | is_baseline | |
Submission ID: 00038 SOSNet, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 134.59 Rank: 55/94 |
0.323 Rank: 75/94 |
0.760 Rank: 64/94 |
0.38654 (±0.00056) Rank: 57/94 |
145.47 Rank: 84/94 |
1271.41 Rank: 69/94 |
4.045 Rank: 74/94 |
0.526 Rank: 69/94 |
0.56072 (±0.00077) Rank: 69/94 |
0.47363 Rank: 63/94 |
Challenge organizers (contact) | sift-def | sosnet (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00582 R2D2Size: 512 bytes. Matches: built-in |
2048.00 | 155.21 Rank: 41/94 |
0.403 Rank: 23/94 |
0.709 Rank: 76/94 |
0.32501 (±0.00085) Rank: 78/94 |
285.11 Rank: 29/94 |
1162.94 Rank: 82/94 |
4.161 Rank: 66/94 |
0.508 Rank: 58/94 |
0.60276 (±0.00062) Rank: 52/94 |
0.46389 Rank: 72/94 |
(contact) | r2d2-5k-p | r2d2-5k-p (128 float32: 512 bytes) | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | ||
Submission ID: 00555 SuperPoint+GIFT+Graph Motion Coh...Size: 512 bytes. Matches: custom |
1940.85 | 281.06 Rank: 16/94 |
0.356 Rank: 31/94 |
0.676 Rank: 80/94 |
0.42283 (±0.00000) Rank: 39/94 |
290.06 Rank: 27/94 |
1887.70 Rank: 19/94 |
4.567 Rank: 22/94 |
0.420 Rank: 14/94 |
0.70395 (±0.00239) Rank: 13/94 |
0.56339 Rank: 21/94 |
Yuan Liu (contact) | superpoint-2k | scale-gift (128 float32: 512 bytes) | Detecting by SuperPoint, descriptors from GIFT, matching by Graph Motion Coherence Network, geometry estimated by DEGENSAC with inlier threshold 1.0 | N/A | N/A | 20-05-10 | is_submission, is_challenge_2020 | |
Submission ID: 00640 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: built-in |
2048.00 | 128.76 Rank: 60/94 |
0.450 Rank: 4/94 |
0.744 Rank: 71/94 |
0.33122 (±0.00042) Rank: 76/94 |
289.47 Rank: 28/94 |
1543.70 Rank: 38/94 |
4.284 Rank: 52/94 |
0.509 Rank: 59/94 |
0.60128 (±0.00311) Rank: 55/94 |
0.46625 Rank: 70/94 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features by score. HardNet ,pre-trained on all datasets, is used as a descriptor head. DEGENSA with degeneracy check on is used as a robust estimator. Cyclic consistency matching with a threshold of 0.95 is used. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-02-17 | is_submission | |
Submission ID: 00039 HardNet, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 138.43 Rank: 54/94 |
0.323 Rank: 75/94 |
0.757 Rank: 67/94 |
0.38490 (±0.00120) Rank: 60/94 |
153.29 Rank: 77/94 |
1306.64 Rank: 62/94 |
4.026 Rank: 76/94 |
0.537 Rank: 75/94 |
0.55573 (±0.00175) Rank: 73/94 |
0.47032 Rank: 67/94 |
Challenge organizers (contact) | sift-def | hardnet (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00516 Key.Net (MS 4-1, Pyramid) + X-Ne...Size: 512 bytes. Matches: built-in |
2039.11 | 265.52 Rank: 18/94 |
0.447 Rank: 13/94 |
0.776 Rank: 51/94 |
0.40559 (±0.00112) Rank: 46/94 |
336.81 Rank: 18/94 |
1559.74 Rank: 36/94 |
4.536 Rank: 24/94 |
0.461 Rank: 32/94 |
0.66406 (±0.00521) Rank: 27/94 |
0.53483 Rank: 36/94 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet-41-pyramidv2 | x-net-lib (128 float32: 512 bytes) | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | ||
Submission ID: 00517 retrained Hardnet with MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 133.39 Rank: 57/94 |
0.333 Rank: 51/94 |
0.808 Rank: 22/94 |
0.40042 (±0.00077) Rank: 49/94 |
210.10 Rank: 47/94 |
1480.67 Rank: 41/94 |
4.226 Rank: 58/94 |
0.479 Rank: 46/94 |
0.63030 (±0.00242) Rank: 41/94 |
0.51536 Rank: 44/94 |
Vu Trung Nghia and Nguyen Trung Hieu (contact) | sift-upright | hardnet (128 float32: 512 bytes) | Using SIFT upright and retrained Hardnet for keypoints and descriptors. For stereo, we use DEGENSAC Chum et al, CVPR05 with optimal settings | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00711 DISK-32D (depth)Size: 128 bytes. Matches: built-in |
2048.00 | 341.83 Rank: 12/94 |
0.448 Rank: 10/94 |
0.842 Rank: 3/94 |
0.47005 (±0.00037) Rank: 11/94 |
441.16 Rank: 10/94 |
2231.55 Rank: 9/94 |
5.524 Rank: 6/94 |
0.415 Rank: 12/94 |
0.70220 (±0.00208) Rank: 14/94 |
0.58613 Rank: 12/94 |
Michal Tyszkiewicz (contact) | disk-2020-09-15-nms-7-depth-32-save-46-imsize-1024-nump-2048 | disk-2020-09-15-nms-7-depth-32-save-46-imsize-1024-nump-2048 (32 float32: 128 bytes) | Local feature model learned via policy gradient, using 32D descriptors. Model trained with a cycle-consistency loss and supervised with depth-based constraints. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 7x7 window, taking the top 2048 features by score. | N/A | N/A | 20-09-16 | is_baseline | |
Submission ID: 00009 GeoDesc (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 116.56 Rank: 66/94 |
0.333 Rank: 51/94 |
0.784 Rank: 41/94 |
0.38660 (±0.00078) Rank: 56/94 |
149.66 Rank: 80/94 |
1237.96 Rank: 71/94 |
4.219 Rank: 61/94 |
0.525 Rank: 68/94 |
0.56687 (±0.00261) Rank: 65/94 |
0.47673 Rank: 60/94 |
Challenge organizers (contact) | sift-def | geodesc-upright (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00623 HardNet-epoch2-OANetSize: 512 bytes. Matches: custom |
2047.76 | 247.45 Rank: 20/94 |
0.347 Rank: 34/94 |
0.765 Rank: 60/94 |
0.44717 (±0.00000) Rank: 25/94 |
254.11 Rank: 39/94 |
1898.45 Rank: 18/94 |
4.503 Rank: 26/94 |
0.437 Rank: 21/94 |
0.67958 (±0.00183) Rank: 21/94 |
0.56337 Rank: 22/94 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift2k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 2k features, using the oanet trained from scratch and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00620 pffNet + SuperPoint + MAGSACSize: 512 bytes. Matches: built-in |
1267.22 | 57.48 Rank: 92/94 |
0.341 Rank: 38/94 |
0.632 Rank: 83/94 |
0.22549 (±0.00077) Rank: 83/94 |
131.91 Rank: 88/94 |
893.88 Rank: 91/94 |
4.339 Rank: 37/94 |
0.531 Rank: 72/94 |
0.55935 (±0.00143) Rank: 71/94 |
0.39242 Rank: 82/94 |
Jongmin Lee, Seungwook Kim, Yoonwoo Jeong (contact) | superpoint | pffnet (128 float32: 512 bytes) | pffNet descriptors + SuperPoint keypoints + MAGSAC outlier-filtering | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00135 CV-DoG-AffNet-HardNet-kornia-PyR...Size: 512 bytes. Matches: built-in |
2047.84 | 105.57 Rank: 77/94 |
0.339 Rank: 40/94 |
0.800 Rank: 30/94 |
0.35887 (±0.00090) Rank: 68/94 |
284.06 Rank: 30/94 |
1788.70 Rank: 24/94 |
4.191 Rank: 63/94 |
0.513 Rank: 62/94 |
0.58536 (±0.00125) Rank: 59/94 |
0.47211 Rank: 64/94 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00026 AKAZE (OpenCV), DEGENSACSize: 61 bytes. Matches: built-in |
1657.83 | 69.77 Rank: 90/94 |
0.355 Rank: 32/94 |
0.622 Rank: 85/94 |
0.17632 (±0.00087) Rank: 87/94 |
224.40 Rank: 45/94 |
795.54 Rank: 93/94 |
3.311 Rank: 88/94 |
0.819 Rank: 88/94 |
0.23020 (±0.00466) Rank: 88/94 |
0.20326 Rank: 88/94 |
Challenge organizers (contact) | akaze-def | akaze (61 uint8: 61 bytes) | AKAZE with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00706 LogPolar-Upright w/ DEGENSAC (no...Size: 512 bytes. Matches: built-in |
1892.71 | 162.20 Rank: 35/94 |
0.333 Rank: 51/94 |
0.807 Rank: 23/94 |
0.45674 (±0.00059) Rank: 19/94 |
211.87 Rank: 46/94 |
1553.36 Rank: 37/94 |
4.331 Rank: 40/94 |
0.471 Rank: 41/94 |
0.63701 (±0.00278) Rank: 37/94 |
0.54688 Rank: 26/94 |
Challenge organizers (contact) | sift-def | logpolar-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV). Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-06-02 | is_baseline | |
Submission ID: 00669 disk_degree(refine)_End-to-EndSize: 512 bytes. Matches: custom |
2048.00 | 591.03 Rank: 1/94 |
0.447 Rank: 15/94 |
0.811 Rank: 18/94 |
0.53291 (±0.00000) Rank: 7/94 |
602.70 Rank: 1/94 |
2633.91 Rank: 1/94 |
5.556 Rank: 4/94 |
0.382 Rank: 6/94 |
0.75413 (±0.00231) Rank: 6/94 |
0.64352 Rank: 7/94 |
Weiyue Zhao (contact) | disk | disk (128 float32: 512 bytes) | disk discriptors, followed by degree(refine)_End-to-End and DEGENSAC. | N/A | N/A | 21-05-07 | is_submission | |
Submission ID: 00022 L2-Net (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 110.72 Rank: 73/94 |
0.333 Rank: 51/94 |
0.801 Rank: 29/94 |
0.40729 (±0.00097) Rank: 45/94 |
171.93 Rank: 64/94 |
1333.56 Rank: 59/94 |
4.226 Rank: 59/94 |
0.497 Rank: 53/94 |
0.58762 (±0.00228) Rank: 57/94 |
0.49746 Rank: 53/94 |
Challenge organizers (contact) | sift-def | l2net-upright (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-23 | is_baseline | |
Submission ID: 00606 HardNet-epoch2-OANetSize: 512 bytes. Matches: custom |
2047.76 | 244.62 Rank: 23/94 |
0.347 Rank: 34/94 |
0.773 Rank: 54/94 |
0.46389 (±0.00000) Rank: 16/94 |
251.11 Rank: 41/94 |
1916.44 Rank: 17/94 |
4.501 Rank: 27/94 |
0.428 Rank: 18/94 |
0.69407 (±0.00216) Rank: 18/94 |
0.57898 Rank: 14/94 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift2k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 2k features, using the oanet and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00013 SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1936.30 | 76.94 Rank: 87/94 |
0.323 Rank: 85/94 |
0.733 Rank: 73/94 |
0.30766 (±0.00075) Rank: 80/94 |
171.92 Rank: 66/94 |
1059.70 Rank: 89/94 |
3.687 Rank: 86/94 |
0.618 Rank: 85/94 |
0.42476 (±0.00103) Rank: 85/94 |
0.36621 Rank: 83/94 |
Challenge organizers (contact) | sift-def | sift (128 float32: 512 bytes) | SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-23 | is_baseline | |
Submission ID: 00513 Personal Inplementation SIFT (GP...Size: 512 bytes. Matches: built-in |
1900.60 | 70.04 Rank: 89/94 |
0.085 Rank: 94/94 |
0.002 Rank: 94/94 |
0.07782 (±0.00022) Rank: 90/94 |
173.06 Rank: 63/94 |
948.78 Rank: 90/94 |
3.816 Rank: 84/94 |
0.887 Rank: 92/94 |
0.09835 (±0.00068) Rank: 91/94 |
0.08809 Rank: 91/94 |
feyman_priv (contact) | sift-gpu | sift-gpu (128 float32: 512 bytes) | SIFT with 2048 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00609 Key.Net-s-ref + HyNet w/ DEGENSA...Size: 512 bytes. Matches: built-in |
2032.93 | 270.04 Rank: 17/94 |
0.461 Rank: 2/94 |
0.820 Rank: 10/94 |
0.46902 (±0.00032) Rank: 12/94 |
280.40 Rank: 33/94 |
1489.62 Rank: 40/94 |
4.689 Rank: 15/94 |
0.436 Rank: 20/94 |
0.68120 (±0.00119) Rank: 20/94 |
0.57511 Rank: 17/94 |
Barroso-Laguna, Axel and Tian, Yurun, Ng, Tony and Mikolajczyk, Krystian (contact) | keynet-s-ref | hynet (128 float32: 512 bytes) | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | ||
Submission ID: 00033 SOSNet, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 119.91 Rank: 62/94 |
0.323 Rank: 75/94 |
0.769 Rank: 58/94 |
0.37632 (±0.00056) Rank: 63/94 |
145.47 Rank: 84/94 |
1271.41 Rank: 69/94 |
4.045 Rank: 74/94 |
0.524 Rank: 67/94 |
0.56072 (±0.00077) Rank: 69/94 |
0.46852 Rank: 69/94 |
Challenge organizers (contact) | sift-def | sosnet (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00701 GeoDesc (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 132.69 Rank: 58/94 |
0.333 Rank: 51/94 |
0.798 Rank: 32/94 |
0.41364 (±0.00033) Rank: 44/94 |
161.05 Rank: 73/94 |
1287.26 Rank: 66/94 |
4.239 Rank: 56/94 |
0.502 Rank: 57/94 |
0.58373 (±0.00402) Rank: 62/94 |
0.49868 Rank: 52/94 |
Challenge organizers (contact) | sift-def | geodesc-upright (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-06-01 | is_baseline | |
Submission ID: 00040 DELF-GLD (128D), DEGENSACSize: 512 bytes. Matches: built-in |
2036.82 | 100.78 Rank: 80/94 |
0.109 Rank: 90/94 |
0.151 Rank: 91/94 |
0.08367 (±0.00047) Rank: 88/94 |
430.68 Rank: 11/94 |
2162.72 Rank: 11/94 |
2.535 Rank: 90/94 |
0.880 Rank: 90/94 |
0.13247 (±0.00116) Rank: 89/94 |
0.10807 Rank: 89/94 |
Challenge organizers (contact) | delf-gld-2k-128d | delf-gld-2k-128d (128 float32: 512 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 128 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with DEGENSAC. | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00564 SIFT2k_2000_HardNet64_train_all_...Size: 512 bytes. Matches: built-in |
1999.83 | 156.93 Rank: 39/94 |
0.336 Rank: 45/94 |
0.814 Rank: 14/94 |
0.42959 (±0.00060) Rank: 35/94 |
244.13 Rank: 43/94 |
1646.76 Rank: 32/94 |
4.334 Rank: 39/94 |
0.459 Rank: 31/94 |
0.64407 (±0.00339) Rank: 32/94 |
0.53683 Rank: 34/94 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | sift2k-2000-hardnet64-train-all-l2-138000 (128 float32: 512 bytes) | SIFT with 2000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 138000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00616 Sift-Fusion_Avg-NM-Net_End-to-En...Size: 512 bytes. Matches: custom |
1892.70 | 142.38 Rank: 50/94 |
0.333 Rank: 67/94 |
0.801 Rank: 28/94 |
0.44818 (±0.00000) Rank: 24/94 |
146.58 Rank: 82/94 |
1339.45 Rank: 58/94 |
4.406 Rank: 34/94 |
0.456 Rank: 28/94 |
0.65563 (±0.00209) Rank: 29/94 |
0.55190 Rank: 24/94 |
Chen Zhao (contact) | siftdef | fusion-avg (128 float32: 512 bytes) | SIFT and Fusion_Avg, followed by NM-Net_End-to-End and DEGENSAC. | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00003 HardNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 141.66 Rank: 51/94 |
0.333 Rank: 51/94 |
0.803 Rank: 27/94 |
0.44235 (±0.00127) Rank: 27/94 |
193.09 Rank: 55/94 |
1435.93 Rank: 47/94 |
4.302 Rank: 48/94 |
0.474 Rank: 44/94 |
0.62275 (±0.00423) Rank: 43/94 |
0.53255 Rank: 37/94 |
Challenge organizers (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00030 Upright Root-SIFT (OpenCV), DEGE...Size: 512 bytes. Matches: built-in |
1892.72 | 111.69 Rank: 72/94 |
0.333 Rank: 47/94 |
0.781 Rank: 46/94 |
0.39547 (±0.00027) Rank: 52/94 |
201.15 Rank: 51/94 |
1347.75 Rank: 56/94 |
4.091 Rank: 68/94 |
0.522 Rank: 66/94 |
0.56078 (±0.00264) Rank: 68/94 |
0.47812 Rank: 59/94 |
Challenge organizers (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Upright Root-SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00027 DELF-GLD (32D), PyRANSACSize: 128 bytes. Matches: built-in |
2036.82 | 91.18 Rank: 84/94 |
0.109 Rank: 90/94 |
0.122 Rank: 92/94 |
0.06197 (±0.00032) Rank: 94/94 |
453.19 Rank: 7/94 |
1805.48 Rank: 21/94 |
2.425 Rank: 92/94 |
0.916 Rank: 93/94 |
0.09163 (±0.00140) Rank: 92/94 |
0.07680 Rank: 93/94 |
Challenge organizers (contact) | delf-gld-2k-32d | delf-gld-2k-32d (32 float32: 128 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 32 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with PyRANSAC (DEGENSAC with the degeneracy check turned off). | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00037 GeoDesc, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 115.17 Rank: 69/94 |
0.323 Rank: 81/94 |
0.742 Rank: 72/94 |
0.34024 (±0.00050) Rank: 73/94 |
120.68 Rank: 91/94 |
1098.70 Rank: 85/94 |
4.002 Rank: 79/94 |
0.551 Rank: 79/94 |
0.50594 (±0.00073) Rank: 82/94 |
0.42309 Rank: 79/94 |
Challenge organizers (contact) | sift-def | geodesc (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-23 | is_baseline | |
Submission ID: 00586 KeyNet+HardNet+NM-Net_v2Size: 512 bytes. Matches: custom |
2042.89 | 160.25 Rank: 37/94 |
0.435 Rank: 18/94 |
0.821 Rank: 7/94 |
0.40264 (±0.00000) Rank: 47/94 |
165.01 Rank: 70/94 |
1134.69 Rank: 84/94 |
4.720 Rank: 14/94 |
0.448 Rank: 25/94 |
0.67488 (±0.00197) Rank: 22/94 |
0.53876 Rank: 30/94 |
Chen Zhao (contact) | keynet | hardnet (128 float32: 512 bytes) | KeyNet and HardNet, followed by NM-Net_v2 and DEGENSAC. | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | |
Submission ID: 00710 DISK-32D (epi)Size: 128 bytes. Matches: built-in |
2048.00 | 352.43 Rank: 10/94 |
0.459 Rank: 3/94 |
0.826 Rank: 5/94 |
0.45946 (±0.00023) Rank: 18/94 |
484.30 Rank: 4/94 |
2324.68 Rank: 5/94 |
5.589 Rank: 3/94 |
0.427 Rank: 16/94 |
0.69938 (±0.00192) Rank: 16/94 |
0.57942 Rank: 13/94 |
Michal Tyszkiewicz (contact) | disk-2020-09-15-nms-5-epi-32-save-49-imsize-1024-nump-2048 | disk-2020-09-15-nms-5-epi-32-save-49-imsize-1024-nump-2048 (32 float32: 128 bytes) | Local feature model learned via policy gradient, using 32D descriptors. Model trained with a cycle-consistency loss and supervised with epipolar constraints. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 5x5 window, taking the top 2048 features by score. | N/A | N/A | 20-09-16 | is_baseline | |
Submission ID: 00642 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: custom |
2048.00 | 140.36 Rank: 52/94 |
0.450 Rank: 4/94 |
0.571 Rank: 87/94 |
0.19708 (±0.00000) Rank: 86/94 |
146.55 Rank: 83/94 |
1279.68 Rank: 67/94 |
4.057 Rank: 70/94 |
0.542 Rank: 78/94 |
0.53330 (±0.00263) Rank: 75/94 |
0.36519 Rank: 84/94 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features by score. HardNet ,pre-trained on all datasets, is used as a descriptor head. Ordered-Aware(OA) Network, trained from scratch on sift-side-8k features (inlier threshold of network output = 1.0, post process method = None, with other settings set to default), is used to compute robust matches and OANet is used in place of robust model estimator. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-03-07 | is_submission | |
Submission ID: 00134 CV-DoG-AffNet-HardNet-kornia-DEG...Size: 512 bytes. Matches: built-in |
2047.84 | 152.12 Rank: 45/94 |
0.339 Rank: 40/94 |
0.791 Rank: 36/94 |
0.41972 (±0.00027) Rank: 41/94 |
284.06 Rank: 30/94 |
1788.70 Rank: 24/94 |
4.191 Rank: 63/94 |
0.511 Rank: 60/94 |
0.58536 (±0.00125) Rank: 59/94 |
0.50254 Rank: 50/94 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00017 ORB (OpenCV), DEGENSACSize: 32 bytes. Matches: built-in |
2031.83 | 63.43 Rank: 91/94 |
0.366 Rank: 29/94 |
0.523 Rank: 89/94 |
0.08243 (±0.00058) Rank: 89/94 |
120.93 Rank: 90/94 |
280.38 Rank: 94/94 |
2.191 Rank: 94/94 |
0.863 Rank: 89/94 |
0.02318 (±0.00229) Rank: 94/94 |
0.05280 Rank: 94/94 |
Challenge organizers (contact) | orb | orb (32 uint8: 32 bytes) | ORB with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00523 SEKDSize: 512 bytes. Matches: built-in |
2043.44 | 116.81 Rank: 65/94 |
0.386 Rank: 26/94 |
0.799 Rank: 31/94 |
0.42811 (±0.00096) Rank: 36/94 |
166.41 Rank: 69/94 |
1163.31 Rank: 81/94 |
4.427 Rank: 31/94 |
0.456 Rank: 29/94 |
0.64919 (±0.00171) Rank: 30/94 |
0.53865 Rank: 31/94 |
Yafei Song, Ling Cai, Mingyang Li (contact) | sekd | sekd (128 float32: 512 bytes) | Our method named SEKD: Self-Evolving Keypoint Detection and Description, where the SEKD model is trained using COCO test images. In this submission each image has up to 2048 SEKD keypoints, and 128-dim float descriptor. We use the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-04-29 | is_submission, is_challenge_2020 | |
Submission ID: 00651 sp_ae_sg_degensacSize: 512 bytes. Matches: custom |
1873.96 | 288.87 Rank: 15/94 |
0.323 Rank: 72/94 |
0.693 Rank: 78/94 |
0.44506 (±0.00000) Rank: 26/94 |
295.52 Rank: 26/94 |
1994.42 Rank: 15/94 |
4.730 Rank: 12/94 |
0.395 Rank: 9/94 |
0.73681 (±0.00306) Rank: 9/94 |
0.59093 Rank: 11/94 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-04-05 | is_submission | |
Submission ID: 00035 HardNet (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 181.81 Rank: 29/94 |
0.333 Rank: 51/94 |
0.781 Rank: 45/94 |
0.43816 (±0.00058) Rank: 30/94 |
193.09 Rank: 55/94 |
1435.93 Rank: 47/94 |
4.302 Rank: 48/94 |
0.473 Rank: 43/94 |
0.62275 (±0.00423) Rank: 43/94 |
0.53045 Rank: 39/94 |
Challenge organizers (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline |
Phototourism: unlimited keypoints, small descriptors (128 bytes)
Note: entries with the same multi-view configuration may seem duplicated. This is normal: performance is averaged across tasks.
Stereo | Multiview | Avg. | ||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Method | NF | NI | Rep. (3 pix.) |
MS (3 pix.) |
mAA (at 10o) |
NM | NL | TL | ATE | mAA (at 100) |
mAA (at 100) |
|||||||||
Submission ID: 00042 ORB (OpenCV), DEGENSACSize: 32 bytes. Matches: built-in |
7150.21 | 161.98 Rank: 2/2 |
0.514 Rank: 2/2 |
0.653 Rank: 2/2 |
0.16159 (±0.00090) Rank: 2/2 |
910.31 Rank: 1/2 |
1423.38 Rank: 2/2 |
2.722 Rank: 2/2 |
0.897 Rank: 2/2 |
0.08054 (±0.00242) Rank: 2/2 |
0.12106 Rank: 2/2 |
Challenge organizers (contact) | orb | orb (32 uint8: 32 bytes) | ORB with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00010 AKAZE (OpenCV), DEGENSACSize: 61 bytes. Matches: built-in |
7857.11 | 246.74 Rank: 1/2 |
0.553 Rank: 1/2 |
0.735 Rank: 1/2 |
0.30717 (±0.00122) Rank: 1/2 |
479.55 Rank: 2/2 |
2778.68 Rank: 1/2 |
3.393 Rank: 1/2 |
0.737 Rank: 1/2 |
0.36048 (±0.00382) Rank: 1/2 |
0.33383 Rank: 1/2 |
Challenge organizers (contact) | akaze-lowth | akaze (61 uint8: 61 bytes) | AKAZE with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline |
Phototourism: restricted keypoints, small descriptors (128 bytes)
Note: entries with the same multi-view configuration may seem duplicated. This is normal: performance is averaged across tasks.
Stereo | Multiview | Avg. | ||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Method | NF | NI | Rep. (3 pix.) |
MS (3 pix.) |
mAA (at 10o) |
NM | NL | TL | ATE | mAA (at 100) |
mAA (at 100) |
|||||||||
Submission ID: 00025 DELF-GLD (32D), DEGENSACSize: 128 bytes. Matches: built-in |
2036.82 | 96.56 Rank: 3/6 |
0.109 Rank: 5/6 |
0.119 Rank: 6/6 |
0.06573 (±0.00030) Rank: 5/6 |
453.19 Rank: 2/6 |
1805.48 Rank: 3/6 |
2.425 Rank: 4/6 |
0.920 Rank: 6/6 |
0.09163 (±0.00140) Rank: 4/6 |
0.07868 Rank: 4/6 |
Challenge organizers (contact) | delf-gld-2k-32d | delf-gld-2k-32d (32 float32: 128 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 32 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with DEGENSAC. | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00711 DISK-32D (depth)Size: 128 bytes. Matches: built-in |
2048.00 | 341.83 Rank: 2/6 |
0.448 Rank: 2/6 |
0.842 Rank: 1/6 |
0.47005 (±0.00037) Rank: 1/6 |
441.16 Rank: 4/6 |
2231.55 Rank: 2/6 |
5.524 Rank: 2/6 |
0.415 Rank: 1/6 |
0.70220 (±0.00208) Rank: 1/6 |
0.58613 Rank: 1/6 |
Michal Tyszkiewicz (contact) | disk-2020-09-15-nms-7-depth-32-save-46-imsize-1024-nump-2048 | disk-2020-09-15-nms-7-depth-32-save-46-imsize-1024-nump-2048 (32 float32: 128 bytes) | Local feature model learned via policy gradient, using 32D descriptors. Model trained with a cycle-consistency loss and supervised with depth-based constraints. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 7x7 window, taking the top 2048 features by score. | N/A | N/A | 20-09-16 | is_baseline | |
Submission ID: 00026 AKAZE (OpenCV), DEGENSACSize: 61 bytes. Matches: built-in |
1657.83 | 69.77 Rank: 5/6 |
0.355 Rank: 4/6 |
0.622 Rank: 3/6 |
0.17632 (±0.00087) Rank: 3/6 |
224.40 Rank: 5/6 |
795.54 Rank: 5/6 |
3.311 Rank: 3/6 |
0.819 Rank: 3/6 |
0.23020 (±0.00466) Rank: 3/6 |
0.20326 Rank: 3/6 |
Challenge organizers (contact) | akaze-def | akaze (61 uint8: 61 bytes) | AKAZE with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00027 DELF-GLD (32D), PyRANSACSize: 128 bytes. Matches: built-in |
2036.82 | 91.18 Rank: 4/6 |
0.109 Rank: 5/6 |
0.122 Rank: 5/6 |
0.06197 (±0.00032) Rank: 6/6 |
453.19 Rank: 2/6 |
1805.48 Rank: 3/6 |
2.425 Rank: 4/6 |
0.916 Rank: 5/6 |
0.09163 (±0.00140) Rank: 4/6 |
0.07680 Rank: 5/6 |
Challenge organizers (contact) | delf-gld-2k-32d | delf-gld-2k-32d (32 float32: 128 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 32 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with PyRANSAC (DEGENSAC with the degeneracy check turned off). | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00710 DISK-32D (epi)Size: 128 bytes. Matches: built-in |
2048.00 | 352.43 Rank: 1/6 |
0.459 Rank: 1/6 |
0.826 Rank: 2/6 |
0.45946 (±0.00023) Rank: 2/6 |
484.30 Rank: 1/6 |
2324.68 Rank: 1/6 |
5.589 Rank: 1/6 |
0.427 Rank: 2/6 |
0.69938 (±0.00192) Rank: 2/6 |
0.57942 Rank: 2/6 |
Michal Tyszkiewicz (contact) | disk-2020-09-15-nms-5-epi-32-save-49-imsize-1024-nump-2048 | disk-2020-09-15-nms-5-epi-32-save-49-imsize-1024-nump-2048 (32 float32: 128 bytes) | Local feature model learned via policy gradient, using 32D descriptors. Model trained with a cycle-consistency loss and supervised with epipolar constraints. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 5x5 window, taking the top 2048 features by score. | N/A | N/A | 20-09-16 | is_baseline | |
Submission ID: 00017 ORB (OpenCV), DEGENSACSize: 32 bytes. Matches: built-in |
2031.83 | 63.43 Rank: 6/6 |
0.366 Rank: 3/6 |
0.523 Rank: 4/6 |
0.08243 (±0.00058) Rank: 4/6 |
120.93 Rank: 6/6 |
280.38 Rank: 6/6 |
2.191 Rank: 6/6 |
0.863 Rank: 4/6 |
0.02318 (±0.00229) Rank: 6/6 |
0.05280 Rank: 6/6 |
Challenge organizers (contact) | orb | orb (32 uint8: 32 bytes) | ORB with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline |
Phototourism: unlimited keypoints, large descriptors (2048 bytes)
Note: entries with the same multi-view configuration may seem duplicated. This is normal: performance is averaged across tasks.
Stereo | Multiview | Avg. | ||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Method | NF | NI | Rep. (3 pix.) |
MS (3 pix.) |
mAA (at 10o) |
NM | NL | TL | ATE | mAA (at 100) |
mAA (at 100) |
|||||||||
Submission ID: 00081 LogPolar w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 591.21 Rank: 33/147 |
0.472 Rank: 103/147 |
0.832 Rank: 83/147 |
0.52385 (±0.00113) Rank: 57/147 |
415.77 Rank: 92/147 |
4054.60 Rank: 72/147 |
4.316 Rank: 84/147 |
0.432 Rank: 68/147 |
0.69284 (±0.00429) Rank: 70/147 |
0.60835 Rank: 63/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00121 CV-DoG-HardNetAmos-8kSize: 512 bytes. Matches: built-in |
7860.98 | 398.65 Rank: 81/147 |
0.472 Rank: 88/147 |
0.863 Rank: 30/147 |
0.53852 (±0.00039) Rank: 44/147 |
356.56 Rank: 109/147 |
3550.63 Rank: 94/147 |
4.275 Rank: 89/147 |
0.439 Rank: 74/147 |
0.68875 (±0.00138) Rank: 76/147 |
0.61364 Rank: 52/147 |
Challenge organizers (contact) | sift8k | hardnetamos (128 float32: 512 bytes) | CV-DoG-HardNetAmos with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00142 CV-DoG-MKD-Concat-pyransacSize: 512 bytes. Matches: built-in |
7860.77 | 208.02 Rank: 136/147 |
0.472 Rank: 82/147 |
0.854 Rank: 34/147 |
0.40612 (±0.00056) Rank: 118/147 |
348.03 Rank: 121/147 |
3507.39 Rank: 100/147 |
4.169 Rank: 112/147 |
0.471 Rank: 113/147 |
0.64763 (±0.00344) Rank: 108/147 |
0.52687 Rank: 117/147 |
Challenge organizers (contact) | sift | mkd-concat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the MKD-Concat descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1811.11147 | N/A | 21-02-05 | is_baseline | |
Submission ID: 00562 HardNet64-train-all-raw64-balanc...Size: 512 bytes. Matches: built-in |
7831.92 | 370.11 Rank: 91/147 |
0.486 Rank: 77/147 |
0.807 Rank: 114/147 |
0.45770 (±0.00060) Rank: 111/147 |
575.43 Rank: 46/147 |
4756.46 Rank: 34/147 |
4.563 Rank: 36/147 |
0.435 Rank: 72/147 |
0.69780 (±0.00191) Rank: 65/147 |
0.57775 Rank: 89/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-raw64-balance-l2-224000-raw64 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 224000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00602 retrained sosnetSize: 512 bytes. Matches: built-in |
7861.11 | 340.97 Rank: 104/147 |
0.472 Rank: 103/147 |
0.840 Rank: 70/147 |
0.48990 (±0.00046) Rank: 88/147 |
623.55 Rank: 34/147 |
4812.36 Rank: 27/147 |
4.233 Rank: 101/147 |
0.458 Rank: 99/147 |
0.66502 (±0.00158) Rank: 95/147 |
0.57746 Rank: 90/147 |
sosnet (contact) | siftdef | sosnet (128 float32: 512 bytes) | default setting and retrained sosnet | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00576 Upright-Sift + X-Net-lib w/ DEGE...Size: 512 bytes. Matches: built-in |
7830.16 | 465.31 Rank: 62/147 |
0.486 Rank: 44/147 |
0.844 Rank: 61/147 |
0.53156 (±0.00027) Rank: 50/147 |
720.29 Rank: 27/147 |
5398.84 Rank: 20/147 |
4.547 Rank: 41/147 |
0.408 Rank: 37/147 |
0.71966 (±0.00160) Rank: 41/147 |
0.62561 Rank: 46/147 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | sift | x-net-lib (128 float32: 512 bytes) | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | ||
Submission ID: 00086 LogPolar-Upright w/ MAGSAC (no F...Size: 512 bytes. Matches: built-in |
7829.63 | 732.28 Rank: 15/147 |
0.486 Rank: 45/147 |
0.844 Rank: 63/147 |
0.54060 (±0.00047) Rank: 42/147 |
505.37 Rank: 70/147 |
4414.11 Rank: 54/147 |
4.518 Rank: 46/147 |
0.421 Rank: 52/147 |
0.71092 (±0.00251) Rank: 46/147 |
0.62576 Rank: 45/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00061 L2-Net-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 358.37 Rank: 96/147 |
0.486 Rank: 45/147 |
0.845 Rank: 56/147 |
0.52012 (±0.00106) Rank: 62/147 |
369.60 Rank: 107/147 |
3538.91 Rank: 97/147 |
4.407 Rank: 60/147 |
0.447 Rank: 89/147 |
0.68108 (±0.00177) Rank: 85/147 |
0.60060 Rank: 72/147 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-23 | is_baseline | |
Submission ID: 00084 LogPolar-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 459.75 Rank: 63/147 |
0.486 Rank: 45/147 |
0.852 Rank: 42/147 |
0.52316 (±0.00034) Rank: 59/147 |
483.93 Rank: 72/147 |
4405.57 Rank: 57/147 |
4.542 Rank: 42/147 |
0.422 Rank: 53/147 |
0.70680 (±0.00247) Rank: 53/147 |
0.61498 Rank: 50/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00638 MKDNet-polarSize: 512 bytes. Matches: built-in |
7862.74 | 337.43 Rank: 107/147 |
0.472 Rank: 91/147 |
0.847 Rank: 51/147 |
0.51376 (±0.00106) Rank: 70/147 |
552.91 Rank: 49/147 |
4786.70 Rank: 29/147 |
4.259 Rank: 96/147 |
0.434 Rank: 71/147 |
0.69907 (±0.00048) Rank: 63/147 |
0.60642 Rank: 69/147 |
(contact) | sift | mkdnet-polar (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-11 | is_submission | |
Submission ID: 00140 CV-DoG-TFeat-kornia-MAGSACSize: 512 bytes. Matches: built-in |
7860.77 | 292.56 Rank: 119/147 |
0.472 Rank: 82/147 |
0.819 Rank: 106/147 |
0.46677 (±0.00109) Rank: 102/147 |
265.53 Rank: 139/147 |
2905.25 Rank: 127/147 |
4.038 Rank: 128/147 |
0.487 Rank: 120/147 |
0.62608 (±0.00138) Rank: 119/147 |
0.54643 Rank: 109/147 |
Challenge organizers (contact) | sift | tfeat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the TFeat descriptor. Implementation: OpenCV + kornia library | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | N/A | 21-02-05 | is_baseline | |
Submission ID: 00634 SIFT8k-HardNetPS-first-submitSize: 512 bytes. Matches: built-in |
7830.09 | 502.15 Rank: 54/147 |
0.486 Rank: 31/147 |
0.841 Rank: 66/147 |
0.49113 (±0.00083) Rank: 85/147 |
651.33 Rank: 29/147 |
5100.71 Rank: 22/147 |
4.370 Rank: 73/147 |
0.423 Rank: 56/147 |
0.70782 (±0.00264) Rank: 51/147 |
0.59948 Rank: 73/147 |
Xudong Zhang, Yuhao Zhou, Huanhuan Fan (contact) | sift | hardnetps (128 float32: 512 bytes) | SIFT and hardnet with 8k features for fisrt subimission. | N/A | N/A | 20-12-15 | is_submission | |
Submission ID: 00601 SIFT8k-HardNet64Size: 512 bytes. Matches: built-in |
6589.88 | 381.80 Rank: 85/147 |
0.467 Rank: 124/147 |
0.852 Rank: 39/147 |
0.52493 (±0.00044) Rank: 55/147 |
420.36 Rank: 89/147 |
3586.82 Rank: 93/147 |
4.494 Rank: 51/147 |
0.429 Rank: 64/147 |
0.70580 (±0.00229) Rank: 55/147 |
0.61536 Rank: 49/147 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, hardnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-05-29 | is_submission, is_challenge_2020 | |
Submission ID: 00591 Guided-Hardnet-Epoch2Size: 512 bytes. Matches: custom |
7829.63 | 761.99 Rank: 13/147 |
0.486 Rank: 45/147 |
0.823 Rank: 102/147 |
0.61081 (±0.00000) Rank: 1/147 |
785.57 Rank: 23/147 |
6330.70 Rank: 7/147 |
4.680 Rank: 7/147 |
0.358 Rank: 2/147 |
0.78288 (±0.00094) Rank: 2/147 |
0.69684 Rank: 1/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 8k features, using the modified guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-24 | is_submission, is_challenge_2020 | |
Submission ID: 00594 CV-DoG-HardNet8-PTSize: 512 bytes. Matches: built-in |
7829.26 | 579.51 Rank: 37/147 |
0.484 Rank: 78/147 |
0.872 Rank: 15/147 |
0.58375 (±0.00046) Rank: 11/147 |
573.22 Rank: 47/147 |
4489.23 Rank: 49/147 |
4.619 Rank: 18/147 |
0.403 Rank: 31/147 |
0.72432 (±0.00357) Rank: 34/147 |
0.65403 Rank: 22/147 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift8k | h8e512pt (128 float32: 512 bytes) | HardNet8 with PCA compression | N/A | N/A | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00611 sift and hardnet64 train scale(1...Size: 512 bytes. Matches: built-in |
7830.09 | 622.13 Rank: 22/147 |
0.486 Rank: 33/147 |
0.871 Rank: 17/147 |
0.58870 (±0.00041) Rank: 5/147 |
899.14 Rank: 14/147 |
6086.16 Rank: 12/147 |
4.647 Rank: 10/147 |
0.386 Rank: 15/147 |
0.75127 (±0.00234) Rank: 14/147 |
0.66999 Rank: 8/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-l2-val-14000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 14000), FLANN disabled | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00597 SIFT8k-HardNet64Size: 512 bytes. Matches: built-in |
6589.88 | 381.75 Rank: 86/147 |
0.467 Rank: 124/147 |
0.852 Rank: 40/147 |
0.52511 (±0.00056) Rank: 54/147 |
420.42 Rank: 88/147 |
3595.93 Rank: 92/147 |
4.496 Rank: 50/147 |
0.417 Rank: 45/147 |
0.70738 (±0.00218) Rank: 52/147 |
0.61625 Rank: 47/147 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, hardnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00636 MKDNet-cartSize: 512 bytes. Matches: built-in |
7862.74 | 260.22 Rank: 127/147 |
0.472 Rank: 91/147 |
0.819 Rank: 107/147 |
0.46478 (±0.00044) Rank: 106/147 |
432.81 Rank: 85/147 |
4142.25 Rank: 70/147 |
4.050 Rank: 127/147 |
0.463 Rank: 104/147 |
0.66048 (±0.00254) Rank: 98/147 |
0.56263 Rank: 102/147 |
(contact) | sift | mkdnet-cart (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-13 | is_submission | |
Submission ID: 00050 GeoDesc-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 410.23 Rank: 76/147 |
0.486 Rank: 45/147 |
0.827 Rank: 95/147 |
0.48259 (±0.00060) Rank: 94/147 |
394.67 Rank: 104/147 |
3863.42 Rank: 81/147 |
4.380 Rank: 67/147 |
0.443 Rank: 81/147 |
0.67682 (±0.00127) Rank: 90/147 |
0.57971 Rank: 87/147 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00034 D2-Net (single-scale), DEGENSACSize: 2048 bytes. Matches: built-in |
5665.31 | 251.55 Rank: 128/147 |
0.288 Rank: 142/147 |
0.360 Rank: 143/147 |
0.18499 (±0.00022) Rank: 136/147 |
740.01 Rank: 26/147 |
4584.07 Rank: 45/147 |
3.506 Rank: 138/147 |
0.696 Rank: 142/147 |
0.41726 (±0.00175) Rank: 141/147 |
0.30112 Rank: 137/147 |
Challenge organizers (contact) | d2net-singlescale | d2net-singlescale (512 float32: 2048 bytes) | D2-Net, single-scale model, up to 8000 features. Trained on the MegaDepth dataset, removing scenes which overlap with the Phototourism test set. Stereo with DEGENSAC and optimal parameters | http://openaccess.thecvf.com/content_CVPR_2019/papers/Dusmanu_D2-Net_A_Trainable_CNN_for_Joint_Description_and_Detection_of_CVPR_2019_paper.pdf | https://github.com/mihaidusmanu/d2-net | 20-04-22 | is_baseline | |
Submission ID: 00076 SOSNet w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 563.27 Rank: 41/147 |
0.472 Rank: 103/147 |
0.846 Rank: 54/147 |
0.55170 (±0.00042) Rank: 32/147 |
508.66 Rank: 67/147 |
4502.19 Rank: 47/147 |
4.377 Rank: 71/147 |
0.405 Rank: 35/147 |
0.71820 (±0.00298) Rank: 42/147 |
0.63495 Rank: 35/147 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00511 r2d2-wasfiSize: 512 bytes. Matches: built-in |
7861.03 | 276.55 Rank: 122/147 |
0.503 Rank: 17/147 |
0.762 Rank: 123/147 |
0.37893 (±0.00023) Rank: 125/147 |
416.35 Rank: 91/147 |
2834.69 Rank: 130/147 |
4.293 Rank: 88/147 |
0.480 Rank: 117/147 |
0.64359 (±0.00103) Rank: 111/147 |
0.51126 Rank: 124/147 |
Chen Shen (contact) | r2d2 | r2d2-wasfi-epoch25-pretrained-n16-8k (128 float32: 512 bytes) | r2d2 with 8k features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00067 GeoDesc w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 453.39 Rank: 65/147 |
0.472 Rank: 103/147 |
0.835 Rank: 79/147 |
0.50559 (±0.00081) Rank: 77/147 |
395.12 Rank: 102/147 |
3838.97 Rank: 83/147 |
4.264 Rank: 94/147 |
0.442 Rank: 79/147 |
0.68032 (±0.00099) Rank: 88/147 |
0.59295 Rank: 82/147 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00079 LogPolar w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 441.77 Rank: 68/147 |
0.472 Rank: 103/147 |
0.852 Rank: 38/147 |
0.53396 (±0.00086) Rank: 48/147 |
415.77 Rank: 92/147 |
4054.60 Rank: 72/147 |
4.316 Rank: 84/147 |
0.432 Rank: 70/147 |
0.69284 (±0.00429) Rank: 70/147 |
0.61340 Rank: 53/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00508 modified-r2d2Size: 512 bytes. Matches: built-in |
7791.82 | 160.64 Rank: 143/147 |
0.517 Rank: 15/147 |
0.785 Rank: 119/147 |
0.32815 (±0.00037) Rank: 127/147 |
328.39 Rank: 130/147 |
2811.27 Rank: 131/147 |
4.250 Rank: 98/147 |
0.488 Rank: 121/147 |
0.62899 (±0.00193) Rank: 115/147 |
0.47857 Rank: 127/147 |
caoliang (contact) | modified-r2d2 | modified-r2d2 (128 float32: 512 bytes) | r2d2 with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | https://europe.naverlabs.com/wp-content/uploads/2019/09/R2D2-Repeatable-and-Reliable-Detector-and-Descriptor-2.pdf | https://github.com/naver/r2d2 | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00514 UprightRootSIFT-AdaLAMSize: 512 bytes. Matches: custom |
6449.42 | 619.87 Rank: 23/147 |
0.436 Rank: 132/147 |
0.758 Rank: 125/147 |
0.31372 (±0.00000) Rank: 129/147 |
640.08 Rank: 32/147 |
5286.27 Rank: 21/147 |
4.343 Rank: 76/147 |
0.398 Rank: 25/147 |
0.72130 (±0.00085) Rank: 40/147 |
0.51751 Rank: 119/147 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Using upright RootSIFT with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. | https://arxiv.org/abs/2006.04250 | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00052 HardNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 363.51 Rank: 93/147 |
0.472 Rank: 103/147 |
0.853 Rank: 37/147 |
0.52325 (±0.00098) Rank: 58/147 |
402.15 Rank: 95/147 |
3895.00 Rank: 79/147 |
4.345 Rank: 74/147 |
0.426 Rank: 61/147 |
0.69653 (±0.00108) Rank: 66/147 |
0.60989 Rank: 60/147 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00501 Guided matching with Upright roo...Size: 512 bytes. Matches: custom |
7829.27 | 609.35 Rank: 27/147 |
0.486 Rank: 29/147 |
0.786 Rank: 118/147 |
0.49072 (±0.00000) Rank: 86/147 |
1363.98 Rank: 3/147 |
7835.75 Rank: 1/147 |
4.219 Rank: 103/147 |
0.454 Rank: 95/147 |
0.68064 (±0.00390) Rank: 87/147 |
0.58568 Rank: 85/147 |
(contact) | sift | upright-root-sift (128 float32: 512 bytes) | In submission | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00122 CV-DoG-HardNetAmos-8kSize: 512 bytes. Matches: built-in |
7860.98 | 528.67 Rank: 45/147 |
0.472 Rank: 88/147 |
0.838 Rank: 72/147 |
0.53291 (±0.00047) Rank: 49/147 |
356.56 Rank: 109/147 |
3550.63 Rank: 94/147 |
4.275 Rank: 89/147 |
0.443 Rank: 82/147 |
0.68875 (±0.00138) Rank: 76/147 |
0.61083 Rank: 58/147 |
Challenge organizers (contact) | sift8k | hardnetamos (128 float32: 512 bytes) | CV-DoG-HardNetAmos with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00558 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 333.94 Rank: 109/147 |
0.284 Rank: 143/147 |
0.427 Rank: 140/147 |
0.01806 (±0.00000) Rank: 145/147 |
341.75 Rank: 127/147 |
2003.56 Rank: 138/147 |
3.735 Rank: 135/147 |
0.593 Rank: 134/147 |
0.51048 (±0.00436) Rank: 134/147 |
0.26427 Rank: 140/147 |
Fabio Bellavia (contact) | sift | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 10) [Mishkin 2019] | N/A | N/A | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00614 ContextDesc Upright + Mutual Che...Size: 512 bytes. Matches: custom |
7830.09 | 647.63 Rank: 18/147 |
0.487 Rank: 25/147 |
0.846 Rank: 55/147 |
0.57826 (±0.00000) Rank: 17/147 |
668.00 Rank: 28/147 |
5612.57 Rank: 17/147 |
4.666 Rank: 8/147 |
0.367 Rank: 4/147 |
0.77041 (±0.00298) Rank: 5/147 |
0.67433 Rank: 5/147 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | contextdesc-upright | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00625 HardNet64-train-all-l2-val-14000...Size: 512 bytes. Matches: built-in |
7830.09 | 605.12 Rank: 29/147 |
0.486 Rank: 33/147 |
0.872 Rank: 16/147 |
0.58776 (±0.00048) Rank: 7/147 |
899.14 Rank: 14/147 |
6095.78 Rank: 11/147 |
4.645 Rank: 11/147 |
0.389 Rank: 19/147 |
0.74849 (±0.00117) Rank: 17/147 |
0.66812 Rank: 11/147 |
Ximin Zheng, Sheng He, Guanlin Liang (contact) | sift8k | hardnet64-train-all-l2-val-14000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 14000), FLANN disabled | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00540 ASLV2+OANetV2+DEGENSACSize: 512 bytes. Matches: custom |
5982.87 | 1044.40 Rank: 5/147 |
0.550 Rank: 11/147 |
0.758 Rank: 124/147 |
0.50173 (±0.00000) Rank: 80/147 |
1067.11 Rank: 6/147 |
4587.47 Rank: 44/147 |
4.948 Rank: 2/147 |
0.386 Rank: 16/147 |
0.75275 (±0.00186) Rank: 13/147 |
0.62724 Rank: 43/147 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | aslv2 | aslv2 (128 float32: 512 bytes) | ASL detector and descriptor, 8k keypoints, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00049 GeoDesc-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 317.71 Rank: 114/147 |
0.486 Rank: 45/147 |
0.850 Rank: 46/147 |
0.48769 (±0.00064) Rank: 90/147 |
394.67 Rank: 104/147 |
3863.42 Rank: 81/147 |
4.380 Rank: 67/147 |
0.440 Rank: 76/147 |
0.67682 (±0.00127) Rank: 90/147 |
0.58225 Rank: 86/147 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00600 ContextDesc Upright + Mutual Che...Size: 512 bytes. Matches: custom |
7830.09 | 624.55 Rank: 20/147 |
0.487 Rank: 25/147 |
0.847 Rank: 53/147 |
0.57344 (±0.00000) Rank: 19/147 |
644.39 Rank: 31/147 |
5427.23 Rank: 19/147 |
4.700 Rank: 4/147 |
0.368 Rank: 5/147 |
0.77043 (±0.00133) Rank: 4/147 |
0.67194 Rank: 6/147 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | contextdesc-upright | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-29 | is_submission, is_challenge_2020 | |
Submission ID: 00532 affnet-hardnetSize: 512 bytes. Matches: built-in |
7925.53 | 226.75 Rank: 133/147 |
0.452 Rank: 130/147 |
0.649 Rank: 133/147 |
0.31958 (±0.00055) Rank: 128/147 |
243.28 Rank: 142/147 |
3155.72 Rank: 119/147 |
3.595 Rank: 137/147 |
0.554 Rank: 133/147 |
0.52578 (±0.00343) Rank: 133/147 |
0.42268 Rank: 130/147 |
caoliang (contact) | affnet | hardnet (128 float32: 512 bytes) | affnet up to 8000 keypoints, harnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1711.06704 | https://github.com/ducha-aiki/affnet | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00071 L2-Net w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 366.00 Rank: 92/147 |
0.472 Rank: 103/147 |
0.852 Rank: 44/147 |
0.52953 (±0.00043) Rank: 51/147 |
339.00 Rank: 128/147 |
3424.94 Rank: 105/147 |
4.206 Rank: 105/147 |
0.452 Rank: 93/147 |
0.66445 (±0.00213) Rank: 96/147 |
0.59699 Rank: 77/147 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-23 | is_baseline | |
Submission ID: 00074 L2-Net w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 481.03 Rank: 57/147 |
0.472 Rank: 103/147 |
0.830 Rank: 91/147 |
0.52519 (±0.00039) Rank: 53/147 |
339.00 Rank: 128/147 |
3424.94 Rank: 105/147 |
4.206 Rank: 105/147 |
0.448 Rank: 90/147 |
0.66445 (±0.00213) Rank: 96/147 |
0.59482 Rank: 81/147 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-23 | is_baseline | |
Submission ID: 00054 HardNet w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 477.64 Rank: 58/147 |
0.472 Rank: 103/147 |
0.830 Rank: 90/147 |
0.51809 (±0.00053) Rank: 64/147 |
402.15 Rank: 95/147 |
3895.00 Rank: 79/147 |
4.345 Rank: 74/147 |
0.423 Rank: 58/147 |
0.69653 (±0.00108) Rank: 66/147 |
0.60731 Rank: 67/147 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00073 L2-Net-Upright w/ MAGSAC (no FLA...Size: 512 bytes. Matches: built-in |
7829.63 | 577.91 Rank: 38/147 |
0.486 Rank: 45/147 |
0.837 Rank: 78/147 |
0.53911 (±0.00062) Rank: 43/147 |
395.53 Rank: 100/147 |
3603.85 Rank: 88/147 |
4.382 Rank: 64/147 |
0.455 Rank: 96/147 |
0.68491 (±0.00338) Rank: 81/147 |
0.61201 Rank: 57/147 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00572 Guided HardNet qhtSize: 512 bytes. Matches: custom |
7829.63 | 885.09 Rank: 8/147 |
0.486 Rank: 45/147 |
0.774 Rank: 122/147 |
0.54663 (±0.00000) Rank: 35/147 |
538.28 Rank: 54/147 |
4687.01 Rank: 36/147 |
4.644 Rank: 13/147 |
0.376 Rank: 10/147 |
0.75721 (±0.00140) Rank: 11/147 |
0.65192 Rank: 25/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-qht (128 float32: 512 bytes) | sift and hardnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | |
Submission ID: 00621 CV-DoG-HardNet8-PTv2Size: 512 bytes. Matches: built-in |
7830.09 | 583.12 Rank: 35/147 |
0.486 Rank: 31/147 |
0.880 Rank: 4/147 |
0.58997 (±0.00054) Rank: 4/147 |
577.37 Rank: 44/147 |
4476.61 Rank: 50/147 |
4.638 Rank: 15/147 |
0.407 Rank: 36/147 |
0.73096 (±0.00227) Rank: 24/147 |
0.66046 Rank: 17/147 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift8k | h8e512pt (128 float32: 512 bytes) | HardNet8 with PCA compression, batch sampling from few images | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00138 CV-DoG-TFeat-kornia-DEGENSACSize: 512 bytes. Matches: built-in |
7860.77 | 234.75 Rank: 132/147 |
0.472 Rank: 82/147 |
0.831 Rank: 86/147 |
0.46486 (±0.00039) Rank: 105/147 |
265.53 Rank: 139/147 |
2905.25 Rank: 127/147 |
4.038 Rank: 128/147 |
0.487 Rank: 119/147 |
0.62608 (±0.00138) Rank: 119/147 |
0.54547 Rank: 111/147 |
Challenge organizers (contact) | sift | tfeat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the TFeat descriptor. Implementation: OpenCV + kornia library | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | N/A | 21-02-05 | is_baseline | |
Submission ID: 00065 SOSNet-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 407.29 Rank: 78/147 |
0.486 Rank: 45/147 |
0.864 Rank: 29/147 |
0.54098 (±0.00086) Rank: 41/147 |
521.37 Rank: 62/147 |
4473.48 Rank: 51/147 |
4.597 Rank: 23/147 |
0.417 Rank: 43/147 |
0.72179 (±0.00330) Rank: 37/147 |
0.63139 Rank: 38/147 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00595 SIFT8k_8000_HardNet64-train-all-...Size: 512 bytes. Matches: built-in |
7830.09 | 600.42 Rank: 31/147 |
0.486 Rank: 33/147 |
0.867 Rank: 23/147 |
0.58265 (±0.00102) Rank: 14/147 |
786.40 Rank: 22/147 |
5595.19 Rank: 18/147 |
4.603 Rank: 22/147 |
0.390 Rank: 20/147 |
0.74372 (±0.00317) Rank: 23/147 |
0.66319 Rank: 16/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-ben86-l2-210000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 210000 and iou 0.86), FLANN disabled | N/A | N/A | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00566 Guided-HardNet2qhtp.jsonSize: 512 bytes. Matches: custom |
7829.63 | 517.72 Rank: 49/147 |
0.486 Rank: 45/147 |
0.874 Rank: 11/147 |
0.58251 (±0.00000) Rank: 15/147 |
448.12 Rank: 81/147 |
4072.82 Rank: 71/147 |
4.607 Rank: 20/147 |
0.387 Rank: 17/147 |
0.74640 (±0.00246) Rank: 20/147 |
0.66445 Rank: 14/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnetnd32qhtp (128 float32: 512 bytes) | sift and hardnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-15 | is_submission, is_challenge_2020 | |
Submission ID: 00080 LogPolar w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 376.42 Rank: 88/147 |
0.472 Rank: 103/147 |
0.837 Rank: 76/147 |
0.50188 (±0.00044) Rank: 79/147 |
399.70 Rank: 97/147 |
4036.37 Rank: 74/147 |
4.340 Rank: 77/147 |
0.429 Rank: 65/147 |
0.69100 (±0.00154) Rank: 73/147 |
0.59644 Rank: 78/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00019 Upright Root-SIFT (OpenCV), DEGE...Size: 512 bytes. Matches: built-in |
7829.24 | 355.69 Rank: 97/147 |
0.487 Rank: 23/147 |
0.844 Rank: 64/147 |
0.51009 (±0.00033) Rank: 74/147 |
548.18 Rank: 51/147 |
4406.41 Rank: 56/147 |
4.385 Rank: 62/147 |
0.442 Rank: 78/147 |
0.68859 (±0.00254) Rank: 79/147 |
0.59934 Rank: 74/147 |
Challenge organizers (contact) | sift-lowth | rootsift-upright (128 float32: 512 bytes) | Upright Root-SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00062 L2-Net-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 471.61 Rank: 60/147 |
0.486 Rank: 45/147 |
0.822 Rank: 103/147 |
0.51450 (±0.00058) Rank: 68/147 |
369.60 Rank: 107/147 |
3538.91 Rank: 97/147 |
4.407 Rank: 60/147 |
0.448 Rank: 91/147 |
0.68108 (±0.00177) Rank: 85/147 |
0.59779 Rank: 76/147 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00628 Multiple kernel local descriptorSize: 512 bytes. Matches: built-in |
7862.74 | 331.20 Rank: 110/147 |
0.472 Rank: 91/147 |
0.829 Rank: 92/147 |
0.48380 (±0.00048) Rank: 93/147 |
541.02 Rank: 53/147 |
4639.37 Rank: 41/147 |
4.144 Rank: 116/147 |
0.463 Rank: 105/147 |
0.66704 (±0.00645) Rank: 93/147 |
0.57542 Rank: 92/147 |
(contact) | sift | mkdpcawt (128 float32: 512 bytes) | based on the paper [Understanding and Improving Kernel Local Descriptors] | https://arxiv.org/pdf/1811.11147 | https://arxiv.org/abs/1811.11147 | 20-10-28 | is_submission | |
Submission ID: 00522 SIFT and HardNet64 train scale(5...Size: 512 bytes. Matches: built-in |
7830.09 | 513.57 Rank: 51/147 |
0.486 Rank: 33/147 |
0.852 Rank: 45/147 |
0.55385 (±0.00076) Rank: 30/147 |
822.07 Rank: 18/147 |
5660.19 Rank: 16/147 |
4.555 Rank: 37/147 |
0.404 Rank: 33/147 |
0.72744 (±0.00455) Rank: 28/147 |
0.64065 Rank: 31/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-raw64 (128 float32: 512 bytes) | SIFT with 8000 keypoints(raw 64), hardnet64 with 128 descriptors, FLANN disabled | N/A | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00530 SIFT-patchSize: 1024 bytes. Matches: built-in |
7861.62 | 83.83 Rank: 146/147 |
0.472 Rank: 99/147 |
0.577 Rank: 136/147 |
0.22977 (±0.00029) Rank: 133/147 |
80.10 Rank: 146/147 |
973.09 Rank: 146/147 |
2.007 Rank: 146/147 |
0.726 Rank: 143/147 |
0.08184 (±0.00372) Rank: 146/147 |
0.15581 Rank: 145/147 |
feyman_priv (contact) | sift8k | l2net-ith-rcface (256 float32: 1024 bytes) | sift8k with l2net trained on google-landmark-dataset-v1(in 1000 class) | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00132 CV-DoG-AffNet-HardNet-kornia-MAG...Size: 512 bytes. Matches: built-in |
7833.97 | 516.85 Rank: 50/147 |
0.486 Rank: 74/147 |
0.870 Rank: 19/147 |
0.54116 (±0.00028) Rank: 40/147 |
580.47 Rank: 40/147 |
4671.35 Rank: 38/147 |
4.565 Rank: 33/147 |
0.403 Rank: 27/147 |
0.72668 (±0.00115) Rank: 30/147 |
0.63392 Rank: 36/147 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00057 HardNet-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 582.52 Rank: 36/147 |
0.486 Rank: 45/147 |
0.841 Rank: 67/147 |
0.53730 (±0.00082) Rank: 46/147 |
480.13 Rank: 74/147 |
4167.34 Rank: 65/147 |
4.566 Rank: 31/147 |
0.416 Rank: 42/147 |
0.71542 (±0.00274) Rank: 44/147 |
0.62636 Rank: 44/147 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00048 GeoDesc w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 362.25 Rank: 94/147 |
0.472 Rank: 103/147 |
0.816 Rank: 110/147 |
0.46887 (±0.00025) Rank: 99/147 |
350.43 Rank: 119/147 |
3601.67 Rank: 90/147 |
4.234 Rank: 99/147 |
0.447 Rank: 88/147 |
0.65553 (±0.00444) Rank: 102/147 |
0.56220 Rank: 103/147 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00120 CV-DoG-HardNetAmos-8kSize: 512 bytes. Matches: built-in |
7860.98 | 265.75 Rank: 125/147 |
0.472 Rank: 88/147 |
0.873 Rank: 13/147 |
0.46073 (±0.00062) Rank: 110/147 |
356.56 Rank: 109/147 |
3550.63 Rank: 94/147 |
4.275 Rank: 89/147 |
0.439 Rank: 75/147 |
0.68875 (±0.00138) Rank: 76/147 |
0.57474 Rank: 93/147 |
Challenge organizers (contact) | sift8k | hardnetamos (128 float32: 512 bytes) | CV-DoG-HardNetAmos with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00502 Guided matching with Upright roo...Size: 512 bytes. Matches: custom |
7829.27 | 491.96 Rank: 56/147 |
0.486 Rank: 29/147 |
0.828 Rank: 94/147 |
0.50310 (±0.01761) Rank: 78/147 |
809.76 Rank: 20/147 |
5966.91 Rank: 13/147 |
4.462 Rank: 54/147 |
0.425 Rank: 60/147 |
0.71009 (±0.00414) Rank: 48/147 |
0.60659 Rank: 68/147 |
(contact) | sift | upright-root-sift (128 float32: 512 bytes) | In submission | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00510 sift and hardnet64Size: 512 bytes. Matches: built-in |
7861.62 | 326.87 Rank: 111/147 |
0.472 Rank: 101/147 |
0.845 Rank: 59/147 |
0.51149 (±0.00018) Rank: 72/147 |
521.58 Rank: 61/147 |
4587.88 Rank: 43/147 |
4.222 Rank: 102/147 |
0.424 Rank: 59/147 |
0.69458 (±0.00340) Rank: 68/147 |
0.60304 Rank: 70/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-v2 (128 float32: 512 bytes) | SIFT with 8000 keypoints, hardnet64 with 128 descriptors | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00106 VL-DoG-SIFT-8kSize: 512 bytes. Matches: built-in |
7880.59 | 326.20 Rank: 112/147 |
0.490 Rank: 20/147 |
0.809 Rank: 113/147 |
0.46326 (±0.00089) Rank: 107/147 |
324.62 Rank: 131/147 |
3030.67 Rank: 123/147 |
4.173 Rank: 109/147 |
0.462 Rank: 103/147 |
0.62829 (±0.00112) Rank: 116/147 |
0.54578 Rank: 110/147 |
Challenge organizers (contact) | dog | vlsift (128 float32: 512 bytes) | VL-DoG-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00618 Sift-Sem + HyNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
7530.17 | 476.84 Rank: 59/147 |
0.490 Rank: 19/147 |
0.882 Rank: 3/147 |
0.57281 (±0.00047) Rank: 21/147 |
809.82 Rank: 19/147 |
5868.33 Rank: 14/147 |
4.585 Rank: 26/147 |
0.390 Rank: 22/147 |
0.74557 (±0.00393) Rank: 21/147 |
0.65919 Rank: 19/147 |
Barroso-Laguna, Axel and Tian, Yurun, Ng, Tony and Mikolajczyk, Krystian (contact) | sift-semantics | hynet (128 float32: 512 bytes) | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | ||
Submission ID: 00068 GeoDesc-Upright w/ DEGENSAC (no ...Size: 512 bytes. Matches: built-in |
7829.63 | 409.87 Rank: 77/147 |
0.486 Rank: 45/147 |
0.869 Rank: 20/147 |
0.52674 (±0.00049) Rank: 52/147 |
458.60 Rank: 77/147 |
4146.83 Rank: 68/147 |
4.412 Rank: 58/147 |
0.431 Rank: 67/147 |
0.70442 (±0.00189) Rank: 57/147 |
0.61558 Rank: 48/147 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00070 GeoDesc-Upright w/ MAGSAC (no FL...Size: 512 bytes. Matches: built-in |
7829.63 | 534.05 Rank: 44/147 |
0.486 Rank: 45/147 |
0.848 Rank: 49/147 |
0.52078 (±0.00028) Rank: 61/147 |
458.60 Rank: 77/147 |
4146.83 Rank: 68/147 |
4.412 Rank: 58/147 |
0.432 Rank: 69/147 |
0.70442 (±0.00189) Rank: 57/147 |
0.61260 Rank: 56/147 |
Challenge organizers (contact) | sift8k | geodesc-upright (128 float32: 512 bytes) | Upright GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00537 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 1476.71 Rank: 1/147 |
0.284 Rank: 143/147 |
0.145 Rank: 147/147 |
0.00178 (±0.00000) Rank: 147/147 |
1506.82 Rank: 2/147 |
2538.44 Rank: 134/147 |
3.388 Rank: 142/147 |
0.669 Rank: 140/147 |
0.42682 (±0.00259) Rank: 140/147 |
0.21430 Rank: 143/147 |
Fabio Bellavia (contact) | sift | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00536 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 1164.84 Rank: 3/147 |
0.368 Rank: 136/147 |
0.341 Rank: 144/147 |
0.00607 (±0.00000) Rank: 146/147 |
1191.31 Rank: 4/147 |
2215.65 Rank: 136/147 |
4.014 Rank: 132/147 |
0.524 Rank: 132/147 |
0.58741 (±0.00250) Rank: 128/147 |
0.29674 Rank: 138/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00075 SOSNet w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 424.61 Rank: 74/147 |
0.472 Rank: 103/147 |
0.868 Rank: 22/147 |
0.55867 (±0.00060) Rank: 26/147 |
508.66 Rank: 67/147 |
4502.19 Rank: 47/147 |
4.377 Rank: 71/147 |
0.404 Rank: 32/147 |
0.71820 (±0.00298) Rank: 42/147 |
0.63844 Rank: 32/147 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-23 | is_baseline | |
Submission ID: 00059 L2-Net w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 297.34 Rank: 118/147 |
0.472 Rank: 103/147 |
0.835 Rank: 81/147 |
0.49523 (±0.00042) Rank: 83/147 |
314.54 Rank: 134/147 |
3314.19 Rank: 114/147 |
4.204 Rank: 107/147 |
0.467 Rank: 110/147 |
0.65665 (±0.00287) Rank: 99/147 |
0.57594 Rank: 91/147 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00042 ORB (OpenCV), DEGENSACSize: 32 bytes. Matches: built-in |
7150.21 | 161.98 Rank: 141/147 |
0.514 Rank: 16/147 |
0.653 Rank: 131/147 |
0.16159 (±0.00090) Rank: 137/147 |
910.31 Rank: 13/147 |
1423.38 Rank: 144/147 |
2.722 Rank: 145/147 |
0.897 Rank: 147/147 |
0.08054 (±0.00242) Rank: 147/147 |
0.12106 Rank: 147/147 |
Challenge organizers (contact) | orb | orb (32 uint8: 32 bytes) | ORB with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00520 SIFT and HardNet64 train scale(5...Size: 512 bytes. Matches: built-in |
7830.09 | 400.19 Rank: 80/147 |
0.486 Rank: 33/147 |
0.837 Rank: 77/147 |
0.50589 (±0.00059) Rank: 76/147 |
619.25 Rank: 35/147 |
4959.78 Rank: 25/147 |
4.384 Rank: 63/147 |
0.456 Rank: 97/147 |
0.69172 (±0.00048) Rank: 72/147 |
0.59880 Rank: 75/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-scale5 (128 float32: 512 bytes) | SIFT with 8000 keypoints(size scaled by 5), hardnet64 with 128 descriptors, FLANN disabled | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00055 HardNet-Upright w/ DEGENSAC (no ...Size: 512 bytes. Matches: built-in |
7829.63 | 527.59 Rank: 47/147 |
0.486 Rank: 45/147 |
0.876 Rank: 7/147 |
0.57279 (±0.00120) Rank: 22/147 |
509.07 Rank: 65/147 |
4250.40 Rank: 60/147 |
4.548 Rank: 39/147 |
0.413 Rank: 40/147 |
0.72309 (±0.00141) Rank: 35/147 |
0.64794 Rank: 27/147 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00613 HardNet64-data-aug-sort-51Size: 512 bytes. Matches: built-in |
7830.09 | 624.09 Rank: 21/147 |
0.486 Rank: 33/147 |
0.870 Rank: 18/147 |
0.58727 (±0.00076) Rank: 8/147 |
964.80 Rank: 9/147 |
6350.68 Rank: 5/147 |
4.644 Rank: 14/147 |
0.383 Rank: 13/147 |
0.74952 (±0.00162) Rank: 16/147 |
0.66839 Rank: 10/147 |
Ximin Zheng, Sheng He, Guanlin Liang (contact) | sift8k | hardnet64-data-aug-sort-51 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 124000 and data augument), FLANN disabled | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00554 ContextDesc-Upright w/ DEGENSAC ...Size: 512 bytes. Matches: built-in |
7830.09 | 470.69 Rank: 61/147 |
0.487 Rank: 25/147 |
0.877 Rank: 5/147 |
0.55697 (±0.00079) Rank: 28/147 |
543.45 Rank: 52/147 |
4543.76 Rank: 46/147 |
4.498 Rank: 49/147 |
0.405 Rank: 34/147 |
0.72902 (±0.00255) Rank: 27/147 |
0.64299 Rank: 30/147 |
Zixin Luo, Jiahui Zhang, Hongkai Chen (contact) | sift-def | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value and use only single orientation to increase performance. FLANN disabled. | https://arxiv.org/abs/1904.04084 | https://github.com/lzx551402/contextdesc | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00571 ContextDesc Upright + OANetV2 + ...Size: 512 bytes. Matches: custom |
7830.09 | 827.45 Rank: 9/147 |
0.487 Rank: 25/147 |
0.775 Rank: 120/147 |
0.53810 (±0.00000) Rank: 45/147 |
855.38 Rank: 16/147 |
6561.52 Rank: 3/147 |
4.522 Rank: 44/147 |
0.370 Rank: 8/147 |
0.76702 (±0.00081) Rank: 6/147 |
0.65256 Rank: 23/147 |
Jiahui Zhang, Zixin Luo, Hongkai Chen (contact) | contextdesc-upright | contextdesc-upright (128 float32: 512 bytes) | ContextDesc with 8000 SIFT features, using improved OANet matcher and DEGENSAC post-processing | N/A | N/A | 20-05-21 | is_submission, is_challenge_2020 | |
Submission ID: 00078 SOSNet-Upright w/ MAGSAC (no FLA...Size: 512 bytes. Matches: built-in |
7829.63 | 679.94 Rank: 17/147 |
0.486 Rank: 45/147 |
0.855 Rank: 33/147 |
0.56626 (±0.00020) Rank: 23/147 |
600.23 Rank: 37/147 |
4765.43 Rank: 32/147 |
4.584 Rank: 27/147 |
0.403 Rank: 28/147 |
0.73027 (±0.00284) Rank: 25/147 |
0.64826 Rank: 26/147 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00630 Superpoint_modifiedSize: 512 bytes. Matches: built-in |
7927.15 | 25.26 Rank: 147/147 |
0.410 Rank: 135/147 |
0.328 Rank: 145/147 |
0.06569 (±0.00036) Rank: 140/147 |
70.01 Rank: 147/147 |
895.58 Rank: 147/147 |
2.963 Rank: 144/147 |
0.849 Rank: 146/147 |
0.18293 (±0.00230) Rank: 144/147 |
0.12431 Rank: 146/147 |
Anonymous (to be released: 2020-6-12) | superpoint-modified | superpoint-modified (128 float32: 512 bytes) | Modified superpoint output. | N/A | N/A | 20-11-27 | is_submission | |
Submission ID: 00063 SOSNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 340.09 Rank: 105/147 |
0.472 Rank: 103/147 |
0.854 Rank: 35/147 |
0.51891 (±0.00067) Rank: 63/147 |
440.21 Rank: 82/147 |
4178.82 Rank: 63/147 |
4.378 Rank: 69/147 |
0.419 Rank: 48/147 |
0.70078 (±0.00375) Rank: 60/147 |
0.60984 Rank: 61/147 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00548 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 343.78 Rank: 102/147 |
0.368 Rank: 136/147 |
0.607 Rank: 135/147 |
0.28795 (±0.00000) Rank: 131/147 |
351.53 Rank: 115/147 |
1699.08 Rank: 141/147 |
4.124 Rank: 120/147 |
0.507 Rank: 126/147 |
0.61370 (±0.00246) Rank: 122/147 |
0.45082 Rank: 129/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.75, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-03 | is_submission, is_challenge_2020 | |
Submission ID: 00521 SIFT8k-giftSize: 512 bytes. Matches: built-in |
6589.48 | 426.77 Rank: 73/147 |
0.467 Rank: 127/147 |
0.830 Rank: 87/147 |
0.48128 (±0.00056) Rank: 95/147 |
452.18 Rank: 79/147 |
3770.17 Rank: 87/147 |
4.595 Rank: 25/147 |
0.444 Rank: 84/147 |
0.67443 (±0.00253) Rank: 92/147 |
0.57785 Rank: 88/147 |
Chen Shen (contact) | sift-def | gift (128 float32: 512 bytes) | sift-nodups, gift with 8k features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00504 ASLFeat (MS) w/ DEGENSAC (no FLA...Size: 512 bytes. Matches: built-in |
6948.83 | 390.79 Rank: 82/147 |
0.550 Rank: 10/147 |
0.797 Rank: 116/147 |
0.46102 (±0.00067) Rank: 109/147 |
510.80 Rank: 64/147 |
3817.14 Rank: 85/147 |
4.577 Rank: 29/147 |
0.444 Rank: 83/147 |
0.68246 (±0.00173) Rank: 84/147 |
0.57174 Rank: 97/147 |
Zixin Luo, Jiahui Zhang (contact) | aslfeat | aslfeat (128 float32: 512 bytes) | ASLFeat (MS) with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. FLANN disabled. | https://arxiv.org/abs/2003.10071 | https://github.com/lzx551402/ASLFeat | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00010 AKAZE (OpenCV), DEGENSACSize: 61 bytes. Matches: built-in |
7857.11 | 246.74 Rank: 130/147 |
0.553 Rank: 9/147 |
0.735 Rank: 127/147 |
0.30717 (±0.00122) Rank: 130/147 |
479.55 Rank: 76/147 |
2778.68 Rank: 132/147 |
3.393 Rank: 141/147 |
0.737 Rank: 145/147 |
0.36048 (±0.00382) Rank: 143/147 |
0.33383 Rank: 132/147 |
Challenge organizers (contact) | akaze-lowth | akaze (61 uint8: 61 bytes) | AKAZE with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00141 CV-DoG-MKD-Concat-DEGENSACSize: 512 bytes. Matches: built-in |
7860.77 | 305.75 Rank: 116/147 |
0.472 Rank: 82/147 |
0.845 Rank: 58/147 |
0.48465 (±0.00122) Rank: 92/147 |
348.03 Rank: 121/147 |
3507.39 Rank: 100/147 |
4.169 Rank: 112/147 |
0.467 Rank: 109/147 |
0.64763 (±0.00344) Rank: 108/147 |
0.56614 Rank: 99/147 |
Challenge organizers (contact) | sift | mkd-concat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the MKD-Concat descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1811.11147 | N/A | 21-02-05 | is_baseline | |
Submission ID: 00051 HardNet w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 432.32 Rank: 72/147 |
0.472 Rank: 103/147 |
0.866 Rank: 25/147 |
0.55430 (±0.00031) Rank: 29/147 |
426.84 Rank: 86/147 |
4001.40 Rank: 77/147 |
4.339 Rank: 79/147 |
0.419 Rank: 47/147 |
0.70962 (±0.00191) Rank: 49/147 |
0.63196 Rank: 37/147 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00639 MKDNet-hardnetSize: 512 bytes. Matches: built-in |
7862.74 | 341.96 Rank: 103/147 |
0.472 Rank: 91/147 |
0.847 Rank: 52/147 |
0.51643 (±0.00099) Rank: 65/147 |
558.95 Rank: 48/147 |
4811.45 Rank: 28/147 |
4.268 Rank: 92/147 |
0.420 Rank: 51/147 |
0.69845 (±0.00564) Rank: 64/147 |
0.60744 Rank: 65/147 |
(contact) | sift | mkdnet-hardnet (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-11 | is_submission | |
Submission ID: 00005 SURF (OpenCV), DEGENSACSize: 256 bytes. Matches: built-in |
7728.57 | 125.94 Rank: 144/147 |
0.432 Rank: 134/147 |
0.684 Rank: 130/147 |
0.24948 (±0.00065) Rank: 132/147 |
928.29 Rank: 12/147 |
3455.28 Rank: 104/147 |
3.428 Rank: 139/147 |
0.733 Rank: 144/147 |
0.40490 (±0.00448) Rank: 142/147 |
0.32719 Rank: 133/147 |
Challenge organizers (contact) | surf-lowth | surf (64 float32: 256 bytes) | SURF with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00069 GeoDesc w/ DEGENSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 348.52 Rank: 101/147 |
0.472 Rank: 103/147 |
0.856 Rank: 32/147 |
0.51112 (±0.00070) Rank: 73/147 |
395.12 Rank: 102/147 |
3838.97 Rank: 83/147 |
4.264 Rank: 94/147 |
0.443 Rank: 80/147 |
0.68032 (±0.00099) Rank: 88/147 |
0.59572 Rank: 80/147 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00559 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 164.43 Rank: 140/147 |
0.284 Rank: 143/147 |
0.650 Rank: 132/147 |
0.05206 (±0.00000) Rank: 142/147 |
167.59 Rank: 144/147 |
1549.91 Rank: 142/147 |
3.989 Rank: 134/147 |
0.622 Rank: 136/147 |
0.47700 (±0.00167) Rank: 136/147 |
0.26453 Rank: 139/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 5) [Mishkin 2019] | N/A | N/A | 20-05-10 | is_submission, is_challenge_2020 | |
Submission ID: 00006 Upright SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
7829.24 | 319.29 Rank: 113/147 |
0.487 Rank: 23/147 |
0.830 Rank: 89/147 |
0.48742 (±0.00042) Rank: 91/147 |
525.57 Rank: 58/147 |
4147.20 Rank: 67/147 |
4.258 Rank: 97/147 |
0.460 Rank: 100/147 |
0.65664 (±0.00231) Rank: 101/147 |
0.57203 Rank: 96/147 |
Challenge organizers (contact) | sift-lowth | sift-upright (128 float32: 512 bytes) | Upright SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00573 Guided-HardNet-qhtSize: 512 bytes. Matches: custom |
7829.63 | 887.83 Rank: 7/147 |
0.486 Rank: 45/147 |
0.774 Rank: 121/147 |
0.55224 (±0.00000) Rank: 31/147 |
538.28 Rank: 54/147 |
4693.85 Rank: 35/147 |
4.650 Rank: 9/147 |
0.381 Rank: 11/147 |
0.75930 (±0.00177) Rank: 8/147 |
0.65577 Rank: 20/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-qht (128 float32: 512 bytes) | sift and hardnet with 8k features, using the modified guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | |
Submission ID: 00066 SOSNet-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 537.85 Rank: 43/147 |
0.486 Rank: 45/147 |
0.841 Rank: 68/147 |
0.53531 (±0.00061) Rank: 47/147 |
521.37 Rank: 62/147 |
4473.48 Rank: 51/147 |
4.597 Rank: 23/147 |
0.414 Rank: 41/147 |
0.72179 (±0.00330) Rank: 37/147 |
0.62855 Rank: 42/147 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00114 VL-Hess-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 348.95 Rank: 100/147 |
0.547 Rank: 12/147 |
0.800 Rank: 115/147 |
0.43347 (±0.00084) Rank: 115/147 |
347.39 Rank: 124/147 |
3209.10 Rank: 116/147 |
4.126 Rank: 117/147 |
0.517 Rank: 128/147 |
0.58657 (±0.00120) Rank: 129/147 |
0.51002 Rank: 125/147 |
Challenge organizers (contact) | hessian | vlsift (128 float32: 512 bytes) | VL-Hess-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00624 Guided-HardNet-OANetSize: 512 bytes. Matches: custom |
7829.63 | 765.34 Rank: 12/147 |
0.486 Rank: 45/147 |
0.820 Rank: 104/147 |
0.60261 (±0.00000) Rank: 2/147 |
788.49 Rank: 21/147 |
6346.62 Rank: 6/147 |
4.682 Rank: 6/147 |
0.355 Rank: 1/147 |
0.78550 (±0.00157) Rank: 1/147 |
0.69405 Rank: 2/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 8k features, first using the oanet trained from scratch then guided-matching, setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00015 Root-SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
7860.73 | 274.85 Rank: 123/147 |
0.472 Rank: 97/147 |
0.845 Rank: 57/147 |
0.48887 (±0.00011) Rank: 89/147 |
437.74 Rank: 84/147 |
3814.80 Rank: 86/147 |
4.151 Rank: 115/147 |
0.458 Rank: 98/147 |
0.65067 (±0.00155) Rank: 104/147 |
0.56977 Rank: 98/147 |
Challenge organizers (contact) | sift-lowth | rootsift (128 float32: 512 bytes) | Root-SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00102 KeyNet-HardNet-8kSize: 512 bytes. Matches: built-in |
7997.63 | 815.40 Rank: 10/147 |
0.582 Rank: 2/147 |
0.788 Rank: 117/147 |
0.47395 (±0.00073) Rank: 97/147 |
356.21 Rank: 112/147 |
3366.01 Rank: 108/147 |
4.319 Rank: 81/147 |
0.464 Rank: 106/147 |
0.64829 (±0.00174) Rank: 105/147 |
0.56112 Rank: 104/147 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00519 SOSNet-Upright-AdaLAMSize: 512 bytes. Matches: custom |
7830.09 | 803.50 Rank: 11/147 |
0.486 Rank: 33/147 |
0.818 Rank: 108/147 |
0.56056 (±0.00000) Rank: 25/147 |
827.85 Rank: 17/147 |
6157.62 Rank: 10/147 |
4.700 Rank: 5/147 |
0.381 Rank: 12/147 |
0.75917 (±0.00446) Rank: 9/147 |
0.65987 Rank: 18/147 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | Using upright SOSNet descriptors with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. Matches post-processed with DEGENSAC. | https://arxiv.org/abs/2006.04250 | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00024 SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
7860.73 | 238.66 Rank: 131/147 |
0.472 Rank: 97/147 |
0.824 Rank: 99/147 |
0.45426 (±0.00097) Rank: 112/147 |
418.86 Rank: 90/147 |
3515.63 Rank: 99/147 |
4.001 Rank: 133/147 |
0.502 Rank: 125/147 |
0.60193 (±0.00185) Rank: 126/147 |
0.52810 Rank: 116/147 |
Challenge organizers (contact) | sift-lowth | sift (128 float32: 512 bytes) | SIFT with (up to) 8000 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00056 HardNet-Upright w/ DEGENSACSize: 512 bytes. Matches: built-in |
7829.63 | 439.36 Rank: 69/147 |
0.486 Rank: 45/147 |
0.864 Rank: 28/147 |
0.54393 (±0.00027) Rank: 39/147 |
480.13 Rank: 74/147 |
4167.34 Rank: 65/147 |
4.566 Rank: 31/147 |
0.417 Rank: 44/147 |
0.71542 (±0.00274) Rank: 44/147 |
0.62967 Rank: 41/147 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00045 D2-Net (multi-scale), DEGENSACSize: 2048 bytes. Matches: built-in |
6924.14 | 338.82 Rank: 106/147 |
0.321 Rank: 141/147 |
0.386 Rank: 141/147 |
0.20486 (±0.00062) Rank: 135/147 |
740.77 Rank: 25/147 |
5696.95 Rank: 15/147 |
3.326 Rank: 143/147 |
0.660 Rank: 139/147 |
0.42782 (±0.00544) Rank: 139/147 |
0.31634 Rank: 134/147 |
Challenge organizers (contact) | d2net-multiscale | d2net-multiscale (512 float32: 2048 bytes) | D2-Net, multi-scale model, up to 8000 features. Trained on the MegaDepth dataset, removing scenes which overlap with the Phototourism test set. Stereo with DEGENSAC and optimal parameters | http://openaccess.thecvf.com/content_CVPR_2019/papers/Dusmanu_D2-Net_A_Trainable_CNN_for_Joint_Description_and_Detection_of_CVPR_2019_paper.pdf | https://github.com/mihaidusmanu/d2-net | 20-04-22 | is_baseline | |
Submission ID: 00047 GeoDesc w/ DEGENSACSize: 512 bytes. Matches: built-in |
7861.11 | 281.36 Rank: 121/147 |
0.472 Rank: 103/147 |
0.838 Rank: 75/147 |
0.47259 (±0.00033) Rank: 98/147 |
350.43 Rank: 119/147 |
3601.67 Rank: 90/147 |
4.234 Rank: 99/147 |
0.447 Rank: 87/147 |
0.65553 (±0.00444) Rank: 102/147 |
0.56406 Rank: 101/147 |
Challenge organizers (contact) | sift8k | geodesc (128 float32: 512 bytes) | GeoDesc descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00072 L2-Net-Upright w/ DEGENSAC (no F...Size: 512 bytes. Matches: built-in |
7829.63 | 435.68 Rank: 70/147 |
0.486 Rank: 45/147 |
0.859 Rank: 31/147 |
0.54497 (±0.00003) Rank: 36/147 |
395.53 Rank: 100/147 |
3603.85 Rank: 88/147 |
4.382 Rank: 64/147 |
0.452 Rank: 94/147 |
0.68491 (±0.00338) Rank: 81/147 |
0.61494 Rank: 51/147 |
Challenge organizers (contact) | sift8k | l2net-upright (128 float32: 512 bytes) | Upright L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00518 UprightRootSIFT-AdaLAMSize: 512 bytes. Matches: custom |
6449.42 | 435.53 Rank: 71/147 |
0.436 Rank: 132/147 |
0.825 Rank: 98/147 |
0.49625 (±0.00000) Rank: 82/147 |
449.49 Rank: 80/147 |
4353.56 Rank: 59/147 |
4.453 Rank: 56/147 |
0.397 Rank: 24/147 |
0.72490 (±0.00190) Rank: 33/147 |
0.61057 Rank: 59/147 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Using upright RootSIFT with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. Matches post-processed with DEGENSAC. | https://arxiv.org/abs/2006.04250 | N/A | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00526 ASLFeat-MSSize: 512 bytes. Matches: built-in |
7384.19 | 456.55 Rank: 64/147 |
0.578 Rank: 5/147 |
0.752 Rank: 126/147 |
0.42208 (±0.00054) Rank: 116/147 |
532.83 Rank: 57/147 |
4002.80 Rank: 76/147 |
4.521 Rank: 45/147 |
0.447 Rank: 86/147 |
0.66568 (±0.00347) Rank: 94/147 |
0.54388 Rank: 113/147 |
Zixin Luo, Jiahui Zhang (contact) | aslfeat-ms | aslfeat-ms (128 float32: 512 bytes) | ASLFeat (joint predition of detectors and descriptors with multi-scale input) with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-04-27 | is_submission, is_challenge_2020 | |
Submission ID: 00627 Multiple kernel local descriptorSize: 512 bytes. Matches: built-in |
7862.74 | 354.57 Rank: 98/147 |
0.472 Rank: 91/147 |
0.845 Rank: 60/147 |
0.50919 (±0.00037) Rank: 75/147 |
580.16 Rank: 43/147 |
4784.46 Rank: 30/147 |
4.210 Rank: 104/147 |
0.436 Rank: 73/147 |
0.68323 (±0.00354) Rank: 83/147 |
0.59621 Rank: 79/147 |
(contact) | sift | mkdlw (128 float32: 512 bytes) | based on the paper [Understanding and Improving Kernel Local Descriptors] | https://arxiv.org/pdf/1811.11147 | https://arxiv.org/abs/1811.11147 | 20-10-28 | is_submission | |
Submission ID: 00143 CV-DoG-MKD-Concat-magsacSize: 512 bytes. Matches: built-in |
7860.77 | 381.37 Rank: 87/147 |
0.472 Rank: 82/147 |
0.833 Rank: 82/147 |
0.48096 (±0.00051) Rank: 96/147 |
348.03 Rank: 121/147 |
3507.39 Rank: 100/147 |
4.169 Rank: 112/147 |
0.471 Rank: 112/147 |
0.64763 (±0.00344) Rank: 108/147 |
0.56429 Rank: 100/147 |
Challenge organizers (contact) | sift | mkd-concat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the MKD-Concat descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1811.11147 | N/A | 21-02-05 | is_baseline | |
Submission ID: 00083 LogPolar-Upright w/ DEGENSAC (no...Size: 512 bytes. Matches: built-in |
7829.63 | 543.18 Rank: 42/147 |
0.486 Rank: 45/147 |
0.865 Rank: 27/147 |
0.55102 (±0.00015) Rank: 33/147 |
505.37 Rank: 70/147 |
4414.11 Rank: 54/147 |
4.518 Rank: 46/147 |
0.422 Rank: 54/147 |
0.71092 (±0.00251) Rank: 46/147 |
0.63097 Rank: 39/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00053 HardNet w/ MAGSAC (no FLANN)Size: 512 bytes. Matches: built-in |
7861.11 | 575.07 Rank: 39/147 |
0.472 Rank: 103/147 |
0.842 Rank: 65/147 |
0.55022 (±0.00001) Rank: 34/147 |
426.84 Rank: 86/147 |
4001.40 Rank: 77/147 |
4.339 Rank: 79/147 |
0.419 Rank: 50/147 |
0.70962 (±0.00191) Rank: 49/147 |
0.62992 Rank: 40/147 |
Challenge organizers (contact) | sift8k | hardnet (128 float32: 512 bytes) | HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00557 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 614.16 Rank: 24/147 |
0.368 Rank: 136/147 |
0.513 Rank: 138/147 |
0.05482 (±0.00000) Rank: 141/147 |
626.57 Rank: 33/147 |
1793.00 Rank: 140/147 |
4.314 Rank: 86/147 |
0.517 Rank: 129/147 |
0.56933 (±0.00268) Rank: 132/147 |
0.31207 Rank: 135/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 10) [Mishkin 2019] | N/A | N/A | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00113 VL-Hess-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 290.24 Rank: 120/147 |
0.547 Rank: 12/147 |
0.814 Rank: 111/147 |
0.44501 (±0.00077) Rank: 114/147 |
347.39 Rank: 124/147 |
3209.10 Rank: 116/147 |
4.126 Rank: 117/147 |
0.519 Rank: 131/147 |
0.58657 (±0.00120) Rank: 129/147 |
0.51579 Rank: 120/147 |
Challenge organizers (contact) | hessian | vlsift (128 float32: 512 bytes) | VL-Hess-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00130 CV-DoG-AffNet-HardNet-kornia-DEG...Size: 512 bytes. Matches: built-in |
7833.97 | 403.41 Rank: 79/147 |
0.486 Rank: 74/147 |
0.883 Rank: 2/147 |
0.54468 (±0.00090) Rank: 37/147 |
580.47 Rank: 40/147 |
4671.35 Rank: 38/147 |
4.565 Rank: 33/147 |
0.402 Rank: 26/147 |
0.72668 (±0.00115) Rank: 30/147 |
0.63568 Rank: 34/147 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00533 upright-sift8k-hardnetSize: 512 bytes. Matches: built-in |
6589.88 | 371.42 Rank: 90/147 |
0.467 Rank: 124/147 |
0.852 Rank: 43/147 |
0.52463 (±0.00070) Rank: 56/147 |
404.02 Rank: 94/147 |
3501.62 Rank: 103/147 |
4.483 Rank: 52/147 |
0.428 Rank: 63/147 |
0.70123 (±0.00081) Rank: 59/147 |
0.61293 Rank: 54/147 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, harnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00541 sift and sosnet64 train scale(12...Size: 512 bytes. Matches: built-in |
7830.09 | 603.56 Rank: 30/147 |
0.486 Rank: 33/147 |
0.866 Rank: 24/147 |
0.58268 (±0.00061) Rank: 13/147 |
957.95 Rank: 10/147 |
6312.38 Rank: 8/147 |
4.607 Rank: 21/147 |
0.385 Rank: 14/147 |
0.74476 (±0.00137) Rank: 22/147 |
0.66372 Rank: 15/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-sos (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 348000), FLANN disabled | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00563 HardNet64-train-all-SOS-812000Size: 512 bytes. Matches: built-in |
7830.09 | 612.51 Rank: 25/147 |
0.486 Rank: 33/147 |
0.869 Rank: 21/147 |
0.58664 (±0.00095) Rank: 9/147 |
972.47 Rank: 8/147 |
6369.03 Rank: 4/147 |
4.612 Rank: 19/147 |
0.388 Rank: 18/147 |
0.74668 (±0.00251) Rank: 18/147 |
0.66666 Rank: 13/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-sos-812000 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 812000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00105 VL-DoG-SIFT-8kSize: 512 bytes. Matches: built-in |
7880.59 | 261.60 Rank: 126/147 |
0.490 Rank: 20/147 |
0.826 Rank: 97/147 |
0.46555 (±0.00055) Rank: 104/147 |
324.62 Rank: 131/147 |
3030.67 Rank: 123/147 |
4.173 Rank: 109/147 |
0.461 Rank: 101/147 |
0.62829 (±0.00112) Rank: 116/147 |
0.54692 Rank: 108/147 |
Challenge organizers (contact) | dog | vlsift (128 float32: 512 bytes) | VL-DoG-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00110 VL-DoGAff-SIFT-8kSize: 512 bytes. Matches: built-in |
7892.05 | 317.13 Rank: 115/147 |
0.482 Rank: 79/147 |
0.820 Rank: 105/147 |
0.46657 (±0.00079) Rank: 103/147 |
311.54 Rank: 136/147 |
3061.54 Rank: 120/147 |
4.105 Rank: 121/147 |
0.475 Rank: 115/147 |
0.62964 (±0.00428) Rank: 112/147 |
0.54810 Rank: 107/147 |
Challenge organizers (contact) | dogaffine | vlsift (128 float32: 512 bytes) | VL-DoGAff-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00598 HarrisZ (more kpt) + DeepOrienta...Size: 512 bytes. Matches: custom |
4478.94 | 563.62 Rank: 40/147 |
0.455 Rank: 128/147 |
0.711 Rank: 128/147 |
0.39767 (±0.00000) Rank: 123/147 |
576.83 Rank: 45/147 |
2707.54 Rank: 133/147 |
4.381 Rank: 66/147 |
0.445 Rank: 85/147 |
0.69090 (±0.00163) Rank: 75/147 |
0.54428 Rank: 112/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ (start scale 2) [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.5, degeneracy check true, confidence 0.98, max iter 500000) [Mishkin 2019] | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00082 LogPolar w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 496.36 Rank: 55/147 |
0.472 Rank: 103/147 |
0.817 Rank: 109/147 |
0.49129 (±0.00011) Rank: 84/147 |
399.70 Rank: 97/147 |
4036.37 Rank: 74/147 |
4.340 Rank: 77/147 |
0.429 Rank: 66/147 |
0.69100 (±0.00154) Rank: 73/147 |
0.59115 Rank: 83/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed (128 float32: 512 bytes) | LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00610 Hardnet-Upright-AdaLAMSize: 512 bytes. Matches: custom |
6556.61 | 627.71 Rank: 19/147 |
0.442 Rank: 131/147 |
0.828 Rank: 93/147 |
0.58300 (±0.00000) Rank: 12/147 |
645.47 Rank: 30/147 |
5074.91 Rank: 23/147 |
4.575 Rank: 30/147 |
0.361 Rank: 3/147 |
0.77056 (±0.00064) Rank: 3/147 |
0.67678 Rank: 4/147 |
Luca Cavalli, Viktor Larsson, Martin Oswald, Torsten Sattler, Marc Pollefeys (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | Using upright Hardnet descriptors with 8000 features, nearest neighbor matching and outlier rejection enforcing local affine consistency within a confidence-based adaptive error tolerance. Matches post-processed with DEGENSAC. | https://arxiv.org/abs/2006.04250 | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00568 Guided-SOSNet-lib-pSize: 512 bytes. Matches: custom |
7829.63 | 508.45 Rank: 52/147 |
0.486 Rank: 45/147 |
0.874 Rank: 12/147 |
0.57982 (±0.00000) Rank: 16/147 |
524.66 Rank: 59/147 |
4618.86 Rank: 42/147 |
4.632 Rank: 16/147 |
0.369 Rank: 6/147 |
0.75888 (±0.00437) Rank: 10/147 |
0.66935 Rank: 9/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhiwei Ruan, Zhongkun Chen, Jingchao Zhou, Pengfei Xu (contact) | sift8k | sosnet-lib-p (128 float32: 512 bytes) | sift and sosnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-15 | is_submission, is_challenge_2020 | |
Submission ID: 00117 VL-HessAffNet-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 299.02 Rank: 117/147 |
0.577 Rank: 6/147 |
0.831 Rank: 84/147 |
0.46793 (±0.00035) Rank: 101/147 |
350.69 Rank: 116/147 |
3327.71 Rank: 111/147 |
4.076 Rank: 124/147 |
0.487 Rank: 118/147 |
0.60691 (±0.00338) Rank: 123/147 |
0.53742 Rank: 114/147 |
Challenge organizers (contact) | hessian | affnetvlsift (128 float32: 512 bytes) | VL-HessAffNet-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00139 CV-DoG-TFeat-kornia-PyRANSACSize: 512 bytes. Matches: built-in |
7860.77 | 160.77 Rank: 142/147 |
0.472 Rank: 82/147 |
0.839 Rank: 71/147 |
0.40079 (±0.00116) Rank: 119/147 |
265.53 Rank: 139/147 |
2905.25 Rank: 127/147 |
4.038 Rank: 128/147 |
0.488 Rank: 122/147 |
0.62608 (±0.00138) Rank: 119/147 |
0.51344 Rank: 123/147 |
Challenge organizers (contact) | sift | tfeat (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by the TFeat descriptor. Implementation: OpenCV + kornia library | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf | 21-02-05 | is_baseline | |
Submission ID: 00567 Guided-HardNet32-v1-lib-qht-pSize: 512 bytes. Matches: custom |
7829.63 | 520.40 Rank: 48/147 |
0.486 Rank: 45/147 |
0.875 Rank: 10/147 |
0.58509 (±0.00000) Rank: 10/147 |
536.93 Rank: 56/147 |
4685.16 Rank: 37/147 |
4.645 Rank: 12/147 |
0.376 Rank: 9/147 |
0.75702 (±0.00159) Rank: 12/147 |
0.67105 Rank: 7/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhiwei Ruan, Zhongkun Chen, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet32-v1-lib-qht-p (128 float32: 512 bytes) | sift and sosnet with 8k features, using the guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-15 | is_submission, is_challenge_2020 | |
Submission ID: 00077 SOSNet-Upright w/ DEGENSAC (no F...Size: 512 bytes. Matches: built-in |
7829.63 | 508.38 Rank: 53/147 |
0.486 Rank: 45/147 |
0.877 Rank: 6/147 |
0.57385 (±0.00041) Rank: 18/147 |
600.23 Rank: 37/147 |
4765.43 Rank: 32/147 |
4.584 Rank: 27/147 |
0.403 Rank: 30/147 |
0.73027 (±0.00284) Rank: 25/147 |
0.65206 Rank: 24/147 |
Challenge organizers (contact) | sift8k | sosnet-upright (128 float32: 512 bytes) | Upright SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-23 | is_baseline | |
Submission ID: 00101 KeyNet-HardNet-8kSize: 512 bytes. Matches: built-in |
7997.63 | 598.28 Rank: 32/147 |
0.582 Rank: 2/147 |
0.826 Rank: 96/147 |
0.49856 (±0.00038) Rank: 81/147 |
356.21 Rank: 112/147 |
3366.01 Rank: 108/147 |
4.319 Rank: 81/147 |
0.464 Rank: 107/147 |
0.64829 (±0.00174) Rank: 105/147 |
0.57342 Rank: 95/147 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00556 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 385.16 Rank: 84/147 |
0.368 Rank: 136/147 |
0.646 Rank: 134/147 |
0.04192 (±0.00000) Rank: 144/147 |
392.52 Rank: 106/147 |
1475.19 Rank: 143/147 |
4.454 Rank: 55/147 |
0.601 Rank: 135/147 |
0.48498 (±0.00118) Rank: 135/147 |
0.26345 Rank: 141/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 5) [Mishkin 2019] | N/A | N/A | 20-05-11 | is_submission, is_challenge_2020 | |
Submission ID: 00524 SIFT-patchSize: 512 bytes. Matches: built-in |
7861.62 | 84.08 Rank: 145/147 |
0.472 Rank: 99/147 |
0.574 Rank: 137/147 |
0.22805 (±0.00076) Rank: 134/147 |
81.93 Rank: 145/147 |
991.42 Rank: 145/147 |
1.934 Rank: 147/147 |
0.687 Rank: 141/147 |
0.08662 (±0.00384) Rank: 145/147 |
0.15733 Rank: 144/147 |
feyman_priv (contact) | sift8k | l2net-arcface (128 float32: 512 bytes) | sift8k with l2net trained on google-landmark-dataset-v1(in 1000 class) | N/A | N/A | 20-04-26 | is_submission, is_challenge_2020 | |
Submission ID: 00546 HarrisZ + DeepOrientation + SOSN...Size: 512 bytes. Matches: custom |
2410.18 | 1020.86 Rank: 6/147 |
0.368 Rank: 136/147 |
0.366 Rank: 142/147 |
0.07780 (±0.00000) Rank: 139/147 |
1047.01 Rank: 7/147 |
2197.09 Rank: 137/147 |
4.023 Rank: 131/147 |
0.509 Rank: 127/147 |
0.59222 (±0.00276) Rank: 127/147 |
0.33501 Rank: 131/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 15, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-02 | is_baseline | |
Submission ID: 00635 MKDNet-indepSize: 512 bytes. Matches: built-in |
7862.74 | 336.51 Rank: 108/147 |
0.472 Rank: 91/147 |
0.848 Rank: 50/147 |
0.51519 (±0.00069) Rank: 67/147 |
550.86 Rank: 50/147 |
4772.22 Rank: 31/147 |
4.265 Rank: 93/147 |
0.427 Rank: 62/147 |
0.70004 (±0.00374) Rank: 62/147 |
0.60762 Rank: 64/147 |
(contact) | sift | mkdnet-indep (128 float32: 512 bytes) | based on the paper [Explicit spatial encoding for deep local descriptors], trained on Liberty set from PhotoTourism dataset | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | https://openaccess.thecvf.com/content_CVPR_2019/papers/Mukundan_Explicit_Spatial_Encoding_for_Deep_Local_Descriptors_CVPR_2019_paper.pdf | 20-12-13 | is_submission | |
Submission ID: 00709 DISK (LCC/depth)Size: 512 bytes. Matches: built-in |
7844.17 | 1238.52 Rank: 2/147 |
0.644 Rank: 1/147 |
0.852 Rank: 41/147 |
0.55847 (±0.00084) Rank: 27/147 |
1663.81 Rank: 1/147 |
7483.98 Rank: 2/147 |
5.922 Rank: 1/147 |
0.391 Rank: 23/147 |
0.75024 (±0.00316) Rank: 15/147 |
0.65435 Rank: 21/147 |
Under review! NeurIPS anonymous submission, ID 1194 (To be released: 21-06-02) | disk-cc-continued-20-imsize-1024-nms-3-nump-8000 | disk-cc-continued-20-imsize-1024-nms-3-nump-8000 (128 float32: 512 bytes) | Local feature model learned via policy gradient. Model trained with a cycle-consistency loss and supervised with depth. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 3x3 window. We take the top 8000 features by score. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020, is_under_review | |
Submission ID: 00535 hardnetlib32P-Upright w/ DEGENSA...Size: 512 bytes. Matches: built-in |
7829.63 | 527.64 Rank: 46/147 |
0.486 Rank: 45/147 |
0.876 Rank: 8/147 |
0.57327 (±0.00078) Rank: 20/147 |
506.73 Rank: 69/147 |
4230.98 Rank: 62/147 |
4.517 Rank: 48/147 |
0.418 Rank: 46/147 |
0.72171 (±0.00156) Rank: 39/147 |
0.64749 Rank: 28/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Jingchao Zhou, Pengfei Xu (contact) | sift8k-no-dups-mps | hardnetlib32P (128 float32: 512 bytes) | sift and hardnet with 8k features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-27 | is_submission, is_challenge_2020 | |
Submission ID: 00108 VL-DoGAff-SIFT-8kSize: 512 bytes. Matches: built-in |
7892.05 | 171.63 Rank: 139/147 |
0.482 Rank: 79/147 |
0.848 Rank: 48/147 |
0.39839 (±0.00122) Rank: 122/147 |
311.54 Rank: 136/147 |
3061.54 Rank: 120/147 |
4.105 Rank: 121/147 |
0.475 Rank: 116/147 |
0.62964 (±0.00428) Rank: 112/147 |
0.51401 Rank: 122/147 |
Challenge organizers (contact) | dogaffine | vlsift (128 float32: 512 bytes) | VL-DoGAff-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00104 VL-DoG-SIFT-8kSize: 512 bytes. Matches: built-in |
7880.59 | 179.68 Rank: 138/147 |
0.490 Rank: 20/147 |
0.835 Rank: 80/147 |
0.39987 (±0.00068) Rank: 120/147 |
324.62 Rank: 131/147 |
3030.67 Rank: 123/147 |
4.173 Rank: 109/147 |
0.467 Rank: 111/147 |
0.62829 (±0.00112) Rank: 116/147 |
0.51408 Rank: 121/147 |
Challenge organizers (contact) | dog | vlsift (128 float32: 512 bytes) | VL-DoG-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00064 SOSNet w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 444.55 Rank: 67/147 |
0.472 Rank: 103/147 |
0.831 Rank: 85/147 |
0.51402 (±0.00057) Rank: 69/147 |
440.21 Rank: 82/147 |
4178.82 Rank: 63/147 |
4.378 Rank: 69/147 |
0.422 Rank: 55/147 |
0.70078 (±0.00375) Rank: 60/147 |
0.60740 Rank: 66/147 |
Challenge organizers (contact) | sift8k | sosnet (128 float32: 512 bytes) | SOSNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00590 Guided-HardNet-epoch4Size: 512 bytes. Matches: custom |
7829.63 | 586.24 Rank: 34/147 |
0.486 Rank: 45/147 |
0.875 Rank: 9/147 |
0.59919 (±0.00000) Rank: 3/147 |
604.81 Rank: 36/147 |
5062.27 Rank: 24/147 |
4.710 Rank: 3/147 |
0.369 Rank: 7/147 |
0.76219 (±0.00253) Rank: 7/147 |
0.68069 Rank: 3/147 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift8k | hardnet-epoch4 (128 float32: 512 bytes) | sift and hardnet with 8k features, using the modified guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-24 | is_submission, is_challenge_2020 | |
Submission ID: 00506 sift and hardnet64Size: 512 bytes. Matches: built-in |
7861.62 | 358.55 Rank: 95/147 |
0.472 Rank: 101/147 |
0.849 Rank: 47/147 |
0.52085 (±0.00009) Rank: 60/147 |
585.33 Rank: 39/147 |
4957.86 Rank: 26/147 |
4.294 Rank: 87/147 |
0.419 Rank: 49/147 |
0.70472 (±0.00257) Rank: 56/147 |
0.61279 Rank: 55/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64 (128 float32: 512 bytes) | SIFT with 8000 keypoints, hardnet64 with 128 descriptors | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00116 VL-HessAffNet-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 209.26 Rank: 135/147 |
0.577 Rank: 6/147 |
0.841 Rank: 69/147 |
0.39330 (±0.00052) Rank: 124/147 |
350.69 Rank: 116/147 |
3327.71 Rank: 111/147 |
4.076 Rank: 124/147 |
0.490 Rank: 124/147 |
0.60691 (±0.00338) Rank: 123/147 |
0.50011 Rank: 126/147 |
Challenge organizers (contact) | hessian | affnetvlsift (128 float32: 512 bytes) | VL-HessAffNet-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00085 LogPolar-Upright w/ MAGSACSize: 512 bytes. Matches: built-in |
7829.63 | 611.19 Rank: 26/147 |
0.486 Rank: 45/147 |
0.830 Rank: 88/147 |
0.51287 (±0.00038) Rank: 71/147 |
483.93 Rank: 72/147 |
4405.57 Rank: 57/147 |
4.542 Rank: 42/147 |
0.423 Rank: 57/147 |
0.70680 (±0.00247) Rank: 53/147 |
0.60983 Rank: 62/147 |
Challenge organizers (contact) | sift8k | logpolar96-fixed-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-23 | is_baseline | |
Submission ID: 00109 VL-DoGAff-SIFT-8kSize: 512 bytes. Matches: built-in |
7892.05 | 250.11 Rank: 129/147 |
0.482 Rank: 79/147 |
0.838 Rank: 74/147 |
0.46795 (±0.00075) Rank: 100/147 |
311.54 Rank: 136/147 |
3061.54 Rank: 120/147 |
4.105 Rank: 121/147 |
0.475 Rank: 114/147 |
0.62964 (±0.00428) Rank: 112/147 |
0.54880 Rank: 106/147 |
Challenge organizers (contact) | dogaffine | vlsift (128 float32: 512 bytes) | VL-DoGAff-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00060 L2-Net w/ MAGSACSize: 512 bytes. Matches: built-in |
7861.11 | 386.81 Rank: 83/147 |
0.472 Rank: 103/147 |
0.813 Rank: 112/147 |
0.49054 (±0.00043) Rank: 87/147 |
314.54 Rank: 134/147 |
3314.19 Rank: 114/147 |
4.204 Rank: 107/147 |
0.465 Rank: 108/147 |
0.65665 (±0.00287) Rank: 99/147 |
0.57360 Rank: 94/147 |
Challenge organizers (contact) | sift8k | l2net (128 float32: 512 bytes) | L2-Net descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. | http://openaccess.thecvf.com/content_cvpr_2017/papers/Tian_L2-Net_Deep_Learning_CVPR_2017_paper.pdf | https://github.com/yuruntian/L2-Net | 20-04-22 | is_baseline | |
Submission ID: 00542 sift and hardnet64 train scale(1...Size: 512 bytes. Matches: built-in |
7830.09 | 607.44 Rank: 28/147 |
0.486 Rank: 33/147 |
0.872 Rank: 14/147 |
0.58801 (±0.00037) Rank: 6/147 |
950.74 Rank: 11/147 |
6286.36 Rank: 9/147 |
4.626 Rank: 17/147 |
0.390 Rank: 21/147 |
0.74662 (±0.00087) Rank: 19/147 |
0.66732 Rank: 12/147 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet64-train-all-l2 (128 float32: 512 bytes) | SIFT with 8000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 138000), FLANN disabled | N/A | N/A | 20-05-01 | is_submission, is_challenge_2020 | |
Submission ID: 00569 Upright-SIFT-HardNetSize: 512 bytes. Matches: built-in |
6667.86 | 376.15 Rank: 89/147 |
0.491 Rank: 18/147 |
0.844 Rank: 62/147 |
0.51560 (±0.00043) Rank: 66/147 |
397.09 Rank: 99/147 |
3417.56 Rank: 107/147 |
4.479 Rank: 53/147 |
0.448 Rank: 92/147 |
0.68803 (±0.00532) Rank: 80/147 |
0.60181 Rank: 71/147 |
caoliang (contact) | sift8k | hardnet (128 float32: 512 bytes) | SIFT up to 8000 keypoints, harnet extract descriptors.Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-05-20 | is_submission, is_challenge_2020 | |
Submission ID: 00118 VL-HessAffNet-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 350.03 Rank: 99/147 |
0.577 Rank: 6/147 |
0.824 Rank: 100/147 |
0.46263 (±0.00079) Rank: 108/147 |
350.69 Rank: 116/147 |
3327.71 Rank: 111/147 |
4.076 Rank: 124/147 |
0.489 Rank: 123/147 |
0.60691 (±0.00338) Rank: 123/147 |
0.53477 Rank: 115/147 |
Challenge organizers (contact) | hessian | affnetvlsift (128 float32: 512 bytes) | VL-HessAffNet-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with MAGSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00058 HardNet-Upright w/ MAGSAC (no FL...Size: 512 bytes. Matches: built-in |
7829.63 | 707.86 Rank: 16/147 |
0.486 Rank: 45/147 |
0.853 Rank: 36/147 |
0.56374 (±0.00069) Rank: 24/147 |
509.07 Rank: 65/147 |
4250.40 Rank: 60/147 |
4.548 Rank: 39/147 |
0.412 Rank: 39/147 |
0.72309 (±0.00141) Rank: 35/147 |
0.64341 Rank: 29/147 |
Challenge organizers (contact) | sift8k | hardnet-upright (128 float32: 512 bytes) | Upright HardNet descriptors on DoG features (OpenCV), extracted with a low detection threshold to generate up to 8000 points. Using the built-in matcher: bidirectional filter with the both strategy. MAGSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00131 CV-DoG-AffNet-HardNet-kornia-PyR...Size: 512 bytes. Matches: built-in |
7833.97 | 267.89 Rank: 124/147 |
0.486 Rank: 74/147 |
0.890 Rank: 1/147 |
0.45050 (±0.00103) Rank: 113/147 |
580.47 Rank: 40/147 |
4671.35 Rank: 38/147 |
4.565 Rank: 33/147 |
0.403 Rank: 29/147 |
0.72668 (±0.00115) Rank: 30/147 |
0.58859 Rank: 84/147 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00549 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 216.23 Rank: 134/147 |
0.284 Rank: 143/147 |
0.456 Rank: 139/147 |
0.14544 (±0.00000) Rank: 138/147 |
221.40 Rank: 143/147 |
1839.87 Rank: 139/147 |
3.608 Rank: 136/147 |
0.641 Rank: 137/147 |
0.47073 (±0.00158) Rank: 137/147 |
0.30808 Rank: 136/147 |
Fabio Bellavia (contact) | sift | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.75, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-03 | is_submission, is_challenge_2020 | |
Submission ID: 00607 HarrisZ (more kpt) + DeepOrienta...Size: 512 bytes. Matches: custom |
4478.94 | 746.02 Rank: 14/147 |
0.455 Rank: 128/147 |
0.699 Rank: 129/147 |
0.40725 (±0.00000) Rank: 117/147 |
763.44 Rank: 24/147 |
2953.35 Rank: 126/147 |
4.414 Rank: 57/147 |
0.442 Rank: 77/147 |
0.69399 (±0.00179) Rank: 69/147 |
0.55062 Rank: 105/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | HarrisZ (start scale 2) [Bellavia et al. 2011] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 0.75, degeneracy check true, confidence 0.98, max iter 500000) [Mishkin 2019] | N/A | N/A | 20-06-03 | is_submission, is_challenge_2020 | |
Submission ID: 00112 VL-Hess-SIFT-8kSize: 512 bytes. Matches: built-in |
8000.00 | 204.37 Rank: 137/147 |
0.547 Rank: 12/147 |
0.823 Rank: 101/147 |
0.36954 (±0.00081) Rank: 126/147 |
347.39 Rank: 124/147 |
3209.10 Rank: 116/147 |
4.126 Rank: 117/147 |
0.517 Rank: 130/147 |
0.58657 (±0.00120) Rank: 129/147 |
0.47806 Rank: 128/147 |
Challenge organizers (contact) | hessian | vlsift (128 float32: 512 bytes) | VL-Hess-SIFT with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline | |
Submission ID: 00545 structured feature descriptionSize: 512 bytes. Matches: built-in |
7830.21 | 411.68 Rank: 75/147 |
0.486 Rank: 43/147 |
0.865 Rank: 26/147 |
0.54442 (±0.00084) Rank: 38/147 |
524.05 Rank: 60/147 |
4467.23 Rank: 53/147 |
4.551 Rank: 38/147 |
0.412 Rank: 38/147 |
0.72729 (±0.00300) Rank: 29/147 |
0.63586 Rank: 33/147 |
Mahdi Abolfazli Esfahani, Han Wang (contact) | sift | sfd (128 float32: 512 bytes) | structured descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC | N/A | N/A | 20-04-29 | is_submission, is_challenge_2020 | |
Submission ID: 00547 SIFT + DeepOrientation + SOSNet ...Size: 512 bytes. Matches: custom |
2826.82 | 1078.85 Rank: 4/147 |
0.284 Rank: 143/147 |
0.166 Rank: 146/147 |
0.04936 (±0.00000) Rank: 143/147 |
1106.68 Rank: 5/147 |
2494.00 Rank: 135/147 |
3.400 Rank: 140/147 |
0.646 Rank: 138/147 |
0.43054 (±0.00185) Rank: 138/147 |
0.23995 Rank: 142/147 |
Fabio Bellavia (contact) | hz | deep-oriented-sosnet (128 float32: 512 bytes) | SIFT (VLFeat implementation) [Lowe 2004] + Deep patch orientation [Yi et al. 2015] + SOSNet [Tian at al. 2019] (weights: sosnet-32x32-hpatches_a.pth) + Blob Matching [unpublished] + Delaunay Triangulation Matching (DTM) [unpublished] + PyRANSAC (threshold 15, degeneracy check true) [Mishkin 2019] | N/A | N/A | 20-05-03 | is_submission, is_challenge_2020 | |
Submission ID: 00100 KeyNet-HardNet-8kSize: 512 bytes. Matches: built-in |
7997.63 | 448.11 Rank: 66/147 |
0.582 Rank: 2/147 |
0.838 Rank: 73/147 |
0.39971 (±0.00147) Rank: 121/147 |
356.21 Rank: 112/147 |
3366.01 Rank: 108/147 |
4.319 Rank: 81/147 |
0.462 Rank: 102/147 |
0.64829 (±0.00174) Rank: 105/147 |
0.52400 Rank: 118/147 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 8000 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-04 | is_baseline |
Phototourism: restricted keypoints, large descriptors (2048 bytes)
Note: entries with the same multi-view configuration may seem duplicated. This is normal: performance is averaged across tasks.
Stereo | Multiview | Avg. | ||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Method | NF | NI | Rep. (3 pix.) |
MS (3 pix.) |
mAA (at 10o) |
NM | NL | TL | ATE | mAA (at 100) |
mAA (at 100) |
|||||||||
Submission ID: 00708 DISK (LCC/depth)Size: 512 bytes. Matches: built-in |
2048.00 | 404.19 Rank: 7/108 |
0.448 Rank: 11/108 |
0.852 Rank: 1/108 |
0.51315 (±0.00028) Rank: 11/108 |
527.48 Rank: 2/108 |
2428.04 Rank: 3/108 |
5.545 Rank: 5/108 |
0.410 Rank: 15/108 |
0.72705 (±0.00094) Rank: 15/108 |
0.62010 Rank: 13/108 |
Michal Tyszkiewicz (contact) | disk-cc-continued-20-imsize-1024-nms-3-nump-2048 | disk-cc-continued-20-imsize-1024-nms-3-nump-2048 (128 float32: 512 bytes) | Local feature model learned via policy gradient. Model trained with a cycle-consistency loss and supervised with depth. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 3x3 window. We take the top 2048 features by score. | N/A | N/A | 20-06-03 | is_submission, is_challenge_2020 | |
Submission ID: 00608 SIFT2k_2048_HardNet64-train-all-...Size: 512 bytes. Matches: custom |
2047.00 | 245.40 Rank: 27/108 |
0.346 Rank: 46/108 |
0.823 Rank: 6/108 |
0.49219 (±0.00000) Rank: 14/108 |
253.37 Rank: 49/108 |
1984.40 Rank: 22/108 |
4.607 Rank: 22/108 |
0.421 Rank: 19/108 |
0.70020 (±0.00213) Rank: 20/108 |
0.59619 Rank: 14/108 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | hardnet64-train-all-l2-138000-matched (128 float32: 512 bytes) | SIFT with 2048 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 138000), FLANN disabled, custom matches | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00702 L2-Net (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 117.11 Rank: 72/108 |
0.333 Rank: 60/108 |
0.808 Rank: 21/108 |
0.41918 (±0.00059) Rank: 46/108 |
179.79 Rank: 70/108 |
1361.34 Rank: 64/108 |
4.232 Rank: 65/108 |
0.481 Rank: 52/108 |
0.59682 (±0.00079) Rank: 61/108 |
0.50800 Rank: 52/108 |
Challenge organizers (contact) | sift-def | l2net-upright (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-06-01 | is_baseline | |
Submission ID: 00663 disk_degree(patch)_End-to-EndSize: 512 bytes. Matches: custom |
2048.00 | 448.22 Rank: 3/108 |
0.447 Rank: 15/108 |
0.834 Rank: 4/108 |
0.53645 (±0.00000) Rank: 9/108 |
458.32 Rank: 5/108 |
2377.71 Rank: 4/108 |
5.657 Rank: 2/108 |
0.373 Rank: 8/108 |
0.75520 (±0.00336) Rank: 8/108 |
0.64582 Rank: 9/108 |
Weiyue Zhao (contact) | disk | disk (128 float32: 512 bytes) | disk discriptors, followed by degree(patch)_End-to-End and DEGENSAC. | N/A | N/A | 21-04-28 | is_submission | |
Submission ID: 00653 sp_ae_sg_degensac_thSize: 512 bytes. Matches: custom |
2048.00 | 293.72 Rank: 18/108 |
0.352 Rank: 42/108 |
0.772 Rank: 59/108 |
0.54380 (±0.00000) Rank: 6/108 |
300.66 Rank: 31/108 |
1801.43 Rank: 29/108 |
4.723 Rank: 18/108 |
0.366 Rank: 4/108 |
0.75988 (±0.00277) Rank: 7/108 |
0.65184 Rank: 7/108 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-04-14 | is_submission | |
Submission ID: 00503 Guided matching hardnetSize: 512 bytes. Matches: custom |
1892.71 | 149.75 Rank: 52/108 |
0.333 Rank: 74/108 |
0.795 Rank: 35/108 |
0.42159 (±0.00000) Rank: 44/108 |
254.13 Rank: 47/108 |
1754.99 Rank: 35/108 |
4.283 Rank: 61/108 |
0.464 Rank: 40/108 |
0.64144 (±0.00207) Rank: 40/108 |
0.53152 Rank: 43/108 |
(contact) | sift | upright-hardnet (128 float32: 512 bytes) | Sift keypoint ; upright hardnet descriptors ; custom matching: use a deep learning based coarse matcher as a first step. In a second step match the keypoint according to descriptor distances but only for the matches that are close to the coarse match prediction | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00703 Upright Root-SIFT (OpenCV), DEGE...Size: 512 bytes. Matches: built-in |
1892.72 | 112.27 Rank: 80/108 |
0.333 Rank: 56/108 |
0.782 Rank: 45/108 |
0.39860 (±0.00077) Rank: 56/108 |
199.34 Rank: 61/108 |
1341.66 Rank: 66/108 |
4.090 Rank: 78/108 |
0.518 Rank: 70/108 |
0.56230 (±0.00234) Rank: 73/108 |
0.48045 Rank: 62/108 |
Challenge organizers (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Upright Root-SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-06-01 | is_baseline | |
Submission ID: 00644 Example: Upright SIFT (OpenCV)Size: 512 bytes. Matches: built-in |
1892.71 | 125.97 Rank: 67/108 |
0.333 Rank: 74/108 |
0.820 Rank: 11/108 |
0.43853 (±0.00030) Rank: 33/108 |
193.06 Rank: 66/108 |
1436.91 Rank: 55/108 |
4.278 Rank: 63/108 |
0.463 Rank: 38/108 |
0.63822 (±0.00469) Rank: 41/108 |
0.53838 Rank: 37/108 |
(contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | SIFT with 2048 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | https://opencv.org | 21-03-13 | is_submission | |
Submission ID: 00028 DELF-GLD (512D), DEGENSACSize: 2048 bytes. Matches: built-in |
2036.82 | 98.87 Rank: 92/108 |
0.109 Rank: 102/108 |
0.170 Rank: 103/108 |
0.09156 (±0.00045) Rank: 99/108 |
437.64 Rank: 11/108 |
2296.72 Rank: 6/108 |
2.567 Rank: 102/108 |
0.877 Rank: 102/108 |
0.14991 (±0.00513) Rank: 101/108 |
0.12073 Rank: 101/108 |
Challenge organizers (contact) | delf-gld-2k-512d | delf-gld-2k-512d (512 float32: 2048 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 512 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with DEGENSAC. | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-23 | is_baseline | |
Submission ID: 00544 Key.Net + X-Net(Lib) w/ DEGENSACSize: 512 bytes. Matches: built-in |
2039.11 | 194.10 Rank: 33/108 |
0.447 Rank: 13/108 |
0.814 Rank: 15/108 |
0.42332 (±0.00096) Rank: 42/108 |
312.61 Rank: 27/108 |
1515.67 Rank: 48/108 |
4.567 Rank: 27/108 |
0.448 Rank: 29/108 |
0.66559 (±0.00233) Rank: 31/108 |
0.54445 Rank: 32/108 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet | x-net-lib (128 float32: 512 bytes) | N/A | N/A | 20-04-28 | is_submission, is_challenge_2020 | ||
Submission ID: 00031 SuperPoint (2k features, NMS=2),...Size: 1024 bytes. Matches: built-in |
2048.00 | 105.52 Rank: 88/108 |
0.376 Rank: 30/108 |
0.624 Rank: 94/108 |
0.27954 (±0.00073) Rank: 90/108 |
130.52 Rank: 102/108 |
1002.03 Rank: 102/108 |
4.211 Rank: 71/108 |
0.593 Rank: 93/108 |
0.51061 (±0.00378) Rank: 90/108 |
0.39508 Rank: 90/108 |
Challenge organizers (contact) | superpoint-nms2-r1200 | superpoint-nms2-r1200 (256 float32: 1024 bytes) | SuperPoint (initial release based on Pytorch). Lowered detection threshold to obtain up to 2k features per image. Using the built-in Non-Maxima Suppression filter at 2 pixels. Images are resized to 1200 pixels on the largest side. | https://arxiv.org/abs/1712.07629 | https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork | 20-04-22 | is_baseline | |
Submission ID: 00603 SuperPoint-128d-masked + SuperGl...Size: 512 bytes. Matches: custom |
2041.09 | 404.73 Rank: 6/108 |
0.387 Rank: 25/108 |
0.774 Rank: 55/108 |
0.56769 (±0.00034) Rank: 3/108 |
415.75 Rank: 17/108 |
2158.03 Rank: 15/108 |
4.932 Rank: 12/108 |
0.357 Rank: 1/108 |
0.76987 (±0.00122) Rank: 3/108 |
0.66878 Rank: 3/108 |
Paul-Edouard Sarlin (contact) | superpoint-k2048-nms3-refine2-r1600forcecubic-masked-d.001 | superpoint-down128 (128 float32: 512 bytes) | SuperPoint detector (2048 keypoints, NMS with radius 3, confidence threshold 0.001, refinement, on 1600-pixel images). Detections on semantic classes sky and people are removed (segmentation from HFNetV2 trained on MIT ADE20K). SuperPoint descriptor, reduced to 128d with a linear autoencoder. SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.2 pixel inlier threshold). | https://arxiv.org/abs/1911.11763 | https://psarlin.com/superglue | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00622 Sift-HardNet-NM-Net_End-to-EndSize: 512 bytes. Matches: custom |
1892.70 | 152.13 Rank: 49/108 |
0.333 Rank: 76/108 |
0.807 Rank: 24/108 |
0.46620 (±0.00000) Rank: 18/108 |
156.54 Rank: 85/108 |
1403.97 Rank: 58/108 |
4.425 Rank: 38/108 |
0.443 Rank: 27/108 |
0.67160 (±0.00184) Rank: 29/108 |
0.56890 Rank: 22/108 |
Chen Zhao (contact) | siftdef | hardnet (128 float32: 512 bytes) | SIFT and HardNet, followed by NM-Net_End-to-End and DEGENSAC. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00668 sp_ae_sg_degensac_plusSize: 512 bytes. Matches: custom |
2048.00 | 394.88 Rank: 8/108 |
0.430 Rank: 19/108 |
0.749 Rank: 75/108 |
0.37510 (±0.00001) Rank: 69/108 |
402.56 Rank: 18/108 |
2246.23 Rank: 9/108 |
5.344 Rank: 7/108 |
0.432 Rank: 24/108 |
0.69832 (±0.00287) Rank: 22/108 |
0.53671 Rank: 40/108 |
(contact) | sid-00668-superpoint | sid-00668-superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-05-03 | is_submission | |
Submission ID: 00667 sp_ae_sg_degensac_xpSize: 512 bytes. Matches: custom |
2048.00 | 446.24 Rank: 4/108 |
0.407 Rank: 22/108 |
0.791 Rank: 38/108 |
0.58918 (±0.00001) Rank: 2/108 |
458.12 Rank: 6/108 |
2290.85 Rank: 8/108 |
5.123 Rank: 8/108 |
0.365 Rank: 3/108 |
0.77685 (±0.00244) Rank: 1/108 |
0.68301 Rank: 1/108 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-04-29 | is_submission | |
Submission ID: 00023 LogPolarDesc, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 116.83 Rank: 73/108 |
0.323 Rank: 84/108 |
0.770 Rank: 60/108 |
0.39335 (±0.00060) Rank: 58/108 |
162.84 Rank: 81/108 |
1385.60 Rank: 62/108 |
4.046 Rank: 81/108 |
0.519 Rank: 71/108 |
0.57149 (±0.00344) Rank: 68/108 |
0.48242 Rank: 60/108 |
Challenge organizers (contact) | sift-def | logpolar (128 float32: 512 bytes) | LogPolarDesc descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-22 | is_baseline | |
Submission ID: 00583 R2D2Size: 512 bytes. Matches: built-in |
2048.00 | 169.38 Rank: 39/108 |
0.398 Rank: 24/108 |
0.695 Rank: 84/108 |
0.33715 (±0.00149) Rank: 80/108 |
306.15 Rank: 28/108 |
1235.83 Rank: 83/108 |
4.156 Rank: 76/108 |
0.496 Rank: 56/108 |
0.60223 (±0.00103) Rank: 58/108 |
0.46969 Rank: 73/108 |
(contact) | r2d2-5k-p-aug | r2d2-5k-p-aug (128 float32: 512 bytes) | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | ||
Submission ID: 00001 HardNet, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 109.52 Rank: 83/108 |
0.323 Rank: 84/108 |
0.776 Rank: 52/108 |
0.38578 (±0.00131) Rank: 64/108 |
153.29 Rank: 89/108 |
1306.64 Rank: 72/108 |
4.026 Rank: 85/108 |
0.539 Rank: 83/108 |
0.55573 (±0.00175) Rank: 79/108 |
0.47075 Rank: 71/108 |
Challenge organizers (contact) | sift-def | hardnet (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline | |
Submission ID: 00585 NM-Net_v2Size: 512 bytes. Matches: custom |
1892.70 | 151.46 Rank: 51/108 |
0.333 Rank: 76/108 |
0.809 Rank: 19/108 |
0.46620 (±0.00000) Rank: 17/108 |
155.91 Rank: 87/108 |
1392.47 Rank: 61/108 |
4.424 Rank: 39/108 |
0.459 Rank: 35/108 |
0.66837 (±0.00125) Rank: 30/108 |
0.56729 Rank: 23/108 |
Chen Zhao (contact) | siftdef | hardnet (128 float32: 512 bytes) | SIFT and HardNet, followed by NM-Net_v2 and DEGENSAC. | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | |
Submission ID: 00589 SuperPoint-128d + SuperGlue + DE...Size: 512 bytes. Matches: custom |
1973.62 | 320.46 Rank: 15/108 |
0.364 Rank: 36/108 |
0.772 Rank: 58/108 |
0.55214 (±0.00098) Rank: 5/108 |
429.49 Rank: 15/108 |
2130.73 Rank: 17/108 |
4.570 Rank: 25/108 |
0.384 Rank: 11/108 |
0.75283 (±0.00086) Rank: 10/108 |
0.65248 Rank: 4/108 |
Paul-Edouard Sarlin (contact) | superpoint-k2048-nms3-refine2-r1600forcecubic | superpoint-down128 (128 float32: 512 bytes) | SuperPoint detector (2048 keypoints, NMS with radius 3, refinement, on 1600-pixel images) and descriptor; reduced to 128d with a linear autoencoder. SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.2 pixels inlier threshold). | https://arxiv.org/abs/1911.11763 | https://psarlin.com/superglue | 20-05-24 | is_submission, is_challenge_2020 | |
Submission ID: 00617 pffNet + SuperPoint + DEGENSACSize: 512 bytes. Matches: built-in |
1267.22 | 55.58 Rank: 107/108 |
0.341 Rank: 47/108 |
0.629 Rank: 92/108 |
0.25514 (±0.00059) Rank: 91/108 |
131.91 Rank: 100/108 |
893.88 Rank: 104/108 |
4.339 Rank: 43/108 |
0.534 Rank: 79/108 |
0.55935 (±0.00143) Rank: 77/108 |
0.40724 Rank: 88/108 |
Jongmin Lee, Seungwook Kim, Yoonwoo Jeong (contact) | superpoint | pffnet (128 float32: 512 bytes) | pffNet descriptors+ SuperPoint keypoints + DEGENSAC outlier-filtering | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00123 KeyNet-HardNet-2kSize: 512 bytes. Matches: built-in |
2048.00 | 134.39 Rank: 62/108 |
0.469 Rank: 1/108 |
0.818 Rank: 12/108 |
0.32725 (±0.00026) Rank: 83/108 |
195.33 Rank: 62/108 |
1276.30 Rank: 78/108 |
4.493 Rank: 34/108 |
0.490 Rank: 55/108 |
0.61606 (±0.00141) Rank: 52/108 |
0.47166 Rank: 70/108 |
Challenge organizers (contact) | keynettuned | vlhardnet (128 float32: 512 bytes) | KeyNet-HardNet with 2048 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with RANSAC | N/A | N/A | 20-05-03 | is_baseline | |
Submission ID: 00008 L2-Net (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 138.44 Rank: 59/108 |
0.333 Rank: 60/108 |
0.786 Rank: 41/108 |
0.39980 (±0.00020) Rank: 55/108 |
171.93 Rank: 73/108 |
1333.56 Rank: 68/108 |
4.226 Rank: 67/108 |
0.496 Rank: 57/108 |
0.58762 (±0.00228) Rank: 62/108 |
0.49371 Rank: 59/108 |
Challenge organizers (contact) | sift-def | l2net-upright (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-22 | is_baseline | |
Submission ID: 00647 sp_ae_sg_ransacSize: 512 bytes. Matches: custom |
1873.96 | 252.53 Rank: 24/108 |
0.323 Rank: 81/108 |
0.663 Rank: 88/108 |
0.06318 (±0.00000) Rank: 106/108 |
256.66 Rank: 45/108 |
1663.09 Rank: 39/108 |
4.771 Rank: 16/108 |
0.563 Rank: 89/108 |
0.52605 (±0.00451) Rank: 86/108 |
0.29461 Rank: 95/108 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-03-29 | is_submission | |
Submission ID: 00646 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: built-in |
2048.00 | 241.83 Rank: 29/108 |
0.450 Rank: 4/108 |
0.689 Rank: 86/108 |
0.32109 (±0.00043) Rank: 85/108 |
301.64 Rank: 29/108 |
1593.08 Rank: 44/108 |
4.285 Rank: 59/108 |
0.497 Rank: 59/108 |
0.60774 (±0.00266) Rank: 54/108 |
0.46442 Rank: 76/108 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features score-wise. HardNet, pre-trained on all datasets, is used as a descriptor head. Graph-Cut(GC)-RANSAC is used as a robust estimator. Cyclic consistency matching with a threshold of 0.95 is used. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-03-28 | is_submission | |
Submission ID: 00509 Upright-Sift + X-Net-lib w/ DEGE...Size: 512 bytes. Matches: built-in |
1892.70 | 116.55 Rank: 76/108 |
0.333 Rank: 80/108 |
0.821 Rank: 9/108 |
0.42382 (±0.00089) Rank: 41/108 |
168.67 Rank: 77/108 |
1319.63 Rank: 70/108 |
4.303 Rank: 55/108 |
0.468 Rank: 45/108 |
0.62953 (±0.00216) Rank: 47/108 |
0.52668 Rank: 46/108 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | sift-def | x-net-lib-upright-no-dups (128 float32: 512 bytes) | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | ||
Submission ID: 00658 disk_degree_end-to-endSize: 512 bytes. Matches: custom |
2048.00 | 346.06 Rank: 12/108 |
0.447 Rank: 15/108 |
0.843 Rank: 2/108 |
0.51061 (±0.00000) Rank: 12/108 |
354.06 Rank: 22/108 |
2170.03 Rank: 12/108 |
5.679 Rank: 1/108 |
0.393 Rank: 12/108 |
0.74934 (±0.00266) Rank: 12/108 |
0.62998 Rank: 12/108 |
Weiyue Zhao (contact) | disk | disk (128 float32: 512 bytes) | disk discriptors, followed by degree_End-to-End and DEGENSAC. | N/A | N/A | 21-04-22 | is_submission | |
Submission ID: 00665 sp_degree(patch)_degensac-1.1Size: 1024 bytes. Matches: custom |
2048.00 | 257.76 Rank: 23/108 |
0.360 Rank: 38/108 |
0.702 Rank: 83/108 |
0.41739 (±0.00000) Rank: 48/108 |
264.72 Rank: 43/108 |
1659.66 Rank: 40/108 |
5.017 Rank: 11/108 |
0.421 Rank: 20/108 |
0.70462 (±0.00152) Rank: 17/108 |
0.56100 Rank: 27/108 |
Weiyue Zhao (contact) | superpoint | superpoint (256 float32: 1024 bytes) | superpoint discriptors, followed by degree(patch)_End-to-End and DEGENSAC th=1.1 | N/A | N/A | 21-04-29 | is_submission | |
Submission ID: 00605 guided-hardnet-epoch2-v2Size: 512 bytes. Matches: custom |
2047.76 | 220.88 Rank: 30/108 |
0.347 Rank: 43/108 |
0.778 Rank: 50/108 |
0.46469 (±0.00000) Rank: 19/108 |
226.84 Rank: 53/108 |
1827.60 Rank: 26/108 |
4.512 Rank: 31/108 |
0.427 Rank: 22/108 |
0.69306 (±0.00160) Rank: 24/108 |
0.57888 Rank: 20/108 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift2k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 2k features, using the oanet and guided matching and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00659 superpoint_degree_end-to-endSize: 1024 bytes. Matches: custom |
2048.00 | 107.59 Rank: 84/108 |
0.360 Rank: 38/108 |
0.797 Rank: 34/108 |
0.02095 (±0.00000) Rank: 108/108 |
111.13 Rank: 108/108 |
862.34 Rank: 106/108 |
5.026 Rank: 10/108 |
0.700 Rank: 97/108 |
0.39130 (±0.00532) Rank: 95/108 |
0.20613 Rank: 98/108 |
Weiyue Zhao (contact) | superpoint | superpoint (256 float32: 1024 bytes) | superpoint discriptors, followed by degree_End-to-End and DEGENSAC. | N/A | N/A | 21-04-23 | is_submission | |
Submission ID: 00645 sp_ae_sgSize: 512 bytes. Matches: custom |
1873.96 | 488.91 Rank: 2/108 |
0.323 Rank: 81/108 |
0.585 Rank: 96/108 |
0.28532 (±0.00000) Rank: 89/108 |
498.60 Rank: 3/108 |
2432.22 Rank: 2/108 |
4.542 Rank: 29/108 |
0.399 Rank: 14/108 |
0.72857 (±0.00361) Rank: 14/108 |
0.50694 Rank: 53/108 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-03-18 | is_submission | |
Submission ID: 00527 SEKDSize: 512 bytes. Matches: built-in |
1786.69 | 106.34 Rank: 86/108 |
0.383 Rank: 29/108 |
0.759 Rank: 71/108 |
0.38886 (±0.00013) Rank: 59/108 |
150.11 Rank: 91/108 |
1063.08 Rank: 100/108 |
4.575 Rank: 23/108 |
0.465 Rank: 42/108 |
0.63303 (±0.00244) Rank: 45/108 |
0.51095 Rank: 51/108 |
Yafei Song, Ling Cai, Mingyang Li (contact) | sekd | sekd (128 float32: 512 bytes) | The name of our method is Self-Evolving Keypoint Detection and Description (SEKD). Now, the SEKD model is trained only using COCO test images. In this submission each image has up to 2048 SEKD keypoints, and 128-dim float descriptor. We use the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-04-27 | is_submission, is_challenge_2020 | |
Submission ID: ????? CV-DoG-HardNet8-PTSize: 512 bytes. Matches: built-in |
2048.00 | 114.84 Rank: 79/108 |
0.321 Rank: 97/108 |
0.790 Rank: 39/108 |
0.38200 (±0.00081) Rank: 66/108 |
117.84 Rank: 107/108 |
1067.13 Rank: 99/108 |
4.051 Rank: 80/108 |
0.531 Rank: 77/108 |
0.53035 (±0.00531) Rank: 84/108 |
0.45617 Rank: 79/108 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift2k | h8e512pt (128 float32: 512 bytes) | [sid:00593] HardNet8 with PCA compression | N/A | N/A | 20-05-27 | is_submission, is_challenge_2020 | |
Submission ID: 00044 Upright SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1892.72 | 104.22 Rank: 90/108 |
0.333 Rank: 56/108 |
0.763 Rank: 68/108 |
0.37078 (±0.00056) Rank: 71/108 |
205.17 Rank: 57/108 |
1300.49 Rank: 74/108 |
4.000 Rank: 91/108 |
0.563 Rank: 88/108 |
0.52549 (±0.00120) Rank: 87/108 |
0.44814 Rank: 81/108 |
Challenge organizers (contact) | sift-def | sift-upright (128 float32: 512 bytes) | Upright SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00004 SOSNet (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 176.54 Rank: 35/108 |
0.333 Rank: 60/108 |
0.782 Rank: 44/108 |
0.43803 (±0.00092) Rank: 35/108 |
183.92 Rank: 67/108 |
1403.66 Rank: 59/108 |
4.314 Rank: 50/108 |
0.474 Rank: 50/108 |
0.62181 (±0.00161) Rank: 50/108 |
0.52992 Rank: 45/108 |
Challenge organizers (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00043 DELF-GLD (128D), PyRANSACSize: 512 bytes. Matches: built-in |
2036.82 | 94.70 Rank: 95/108 |
0.109 Rank: 102/108 |
0.156 Rank: 104/108 |
0.07298 (±0.00056) Rank: 104/108 |
430.68 Rank: 13/108 |
2162.72 Rank: 13/108 |
2.535 Rank: 104/108 |
0.887 Rank: 105/108 |
0.13247 (±0.00116) Rank: 103/108 |
0.10272 Rank: 104/108 |
Challenge organizers (contact) | delf-gld-2k-128d | delf-gld-2k-128d (128 float32: 512 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 128 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with PyRANSAC (DEGENSAC with the degeneracy check turned off). | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-23 | is_baseline | |
Submission ID: 00612 SuperPoint-128d-adapt + SuperGlu...Size: 512 bytes. Matches: custom |
2048.00 | 441.49 Rank: 5/108 |
0.407 Rank: 21/108 |
0.789 Rank: 40/108 |
0.59034 (±0.00050) Rank: 1/108 |
452.99 Rank: 9/108 |
2245.43 Rank: 10/108 |
5.092 Rank: 9/108 |
0.358 Rank: 2/108 |
0.77337 (±0.00167) Rank: 2/108 |
0.68186 Rank: 2/108 |
Paul-Edouard Sarlin (contact) | superpoint-k2048-nms4-refine2-r1600forcecubic-masked-d.001-adapt50 | superpoint-down128 (128 float32: 512 bytes) | SuperPoint detector (2048 keypoints, NMS with radius 4, confidence threshold 0.001, refinement, on 1600-pixel images). The detection heatmap is improved with test-time homographic adaptation (50 iterations), and detections on semantic classes sky and people are removed (segmentation from HFNetV2 trained on MIT ADE20K). SuperPoint descriptor, reduced to 128d with a linear autoencoder. SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.1 pixel inlier threshold). | https://arxiv.org/abs/1911.11763 | https://psarlin.com/superglue | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00615 SIFT-Fusion_Max-NM-Net_End-to-En...Size: 512 bytes. Matches: custom |
1892.70 | 115.46 Rank: 77/108 |
0.333 Rank: 76/108 |
0.783 Rank: 43/108 |
0.40244 (±0.00000) Rank: 53/108 |
119.23 Rank: 106/108 |
1167.54 Rank: 91/108 |
4.367 Rank: 41/108 |
0.489 Rank: 54/108 |
0.60720 (±0.00114) Rank: 55/108 |
0.50482 Rank: 54/108 |
Chen Zhao (contact) | siftdef | fusion-max (128 float32: 512 bytes) | SIFT and Fusion_Max, followed by NM-Net_End-to-End and DEGENSAC. | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00002 SURF (OpenCV), DEGENSACSize: 256 bytes. Matches: built-in |
2010.88 | 71.90 Rank: 102/108 |
0.321 Rank: 96/108 |
0.634 Rank: 89/108 |
0.20761 (±0.00029) Rank: 94/108 |
300.02 Rank: 32/108 |
1208.06 Rank: 88/108 |
3.549 Rank: 97/108 |
0.695 Rank: 96/108 |
0.37845 (±0.00399) Rank: 97/108 |
0.29303 Rank: 96/108 |
Challenge organizers (contact) | surf-def | surf (64 float32: 256 bytes) | SURF with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00641 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: built-in |
2048.00 | 219.72 Rank: 31/108 |
0.450 Rank: 4/108 |
0.722 Rank: 81/108 |
0.34184 (±0.00032) Rank: 78/108 |
301.64 Rank: 29/108 |
1595.30 Rank: 43/108 |
4.278 Rank: 62/108 |
0.499 Rank: 60/108 |
0.60963 (±0.00152) Rank: 53/108 |
0.47574 Rank: 66/108 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features by score. HardNet ,pre-trained on all datasets, is used as a descriptor head. MAGSAC is used as a robust estimator. Cyclic consistency matching with a threshold of 0.95 is used. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-03-04 | is_submission | |
Submission ID: 00705 Upright SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1892.72 | 104.74 Rank: 89/108 |
0.333 Rank: 56/108 |
0.764 Rank: 67/108 |
0.37292 (±0.00081) Rank: 70/108 |
204.13 Rank: 58/108 |
1293.75 Rank: 75/108 |
4.008 Rank: 88/108 |
0.567 Rank: 91/108 |
0.52808 (±0.00416) Rank: 85/108 |
0.45050 Rank: 80/108 |
Challenge organizers (contact) | sift-def | sift-upright (128 float32: 512 bytes) | Upright SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-06-01 | is_baseline | |
Submission ID: 00014 L2-Net, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 107.38 Rank: 85/108 |
0.323 Rank: 90/108 |
0.766 Rank: 64/108 |
0.35649 (±0.00110) Rank: 75/108 |
138.97 Rank: 98/108 |
1211.49 Rank: 85/108 |
3.948 Rank: 92/108 |
0.538 Rank: 82/108 |
0.52019 (±0.00120) Rank: 88/108 |
0.43834 Rank: 83/108 |
Challenge organizers (contact) | sift-def | l2net (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-23 | is_baseline | |
Submission ID: 00507 sift and hardnet512Size: 2048 bytes. Matches: built-in |
2000.00 | 98.07 Rank: 93/108 |
0.321 Rank: 98/108 |
0.778 Rank: 49/108 |
0.35047 (±0.00046) Rank: 76/108 |
153.92 Rank: 88/108 |
1310.53 Rank: 71/108 |
4.017 Rank: 87/108 |
0.515 Rank: 68/108 |
0.56598 (±0.00185) Rank: 72/108 |
0.45822 Rank: 78/108 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift8k | hardnet512 (512 float32: 2048 bytes) | SIFT with 2000 keypoints, hardnet512 with 512 descriptors | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00032 Root-SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1936.30 | 83.46 Rank: 100/108 |
0.323 Rank: 94/108 |
0.756 Rank: 73/108 |
0.33605 (±0.00092) Rank: 81/108 |
167.84 Rank: 78/108 |
1141.88 Rank: 94/108 |
3.788 Rank: 95/108 |
0.576 Rank: 92/108 |
0.46826 (±0.00154) Rank: 93/108 |
0.40216 Rank: 89/108 |
Challenge organizers (contact) | sift-def | rootsift (128 float32: 512 bytes) | Root-SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00041 SuperPoint (2k features, NMS=3),...Size: 1024 bytes. Matches: built-in |
2048.00 | 117.13 Rank: 71/108 |
0.368 Rank: 31/108 |
0.628 Rank: 93/108 |
0.28679 (±0.00056) Rank: 88/108 |
156.04 Rank: 86/108 |
1121.20 Rank: 96/108 |
4.310 Rank: 54/108 |
0.566 Rank: 90/108 |
0.53829 (±0.00021) Rank: 82/108 |
0.41254 Rank: 87/108 |
Challenge organizers (contact) | superpoint-nms3-r1200 | superpoint-nms3-r1200 (256 float32: 1024 bytes) | SuperPoint (initial release based on Pytorch). Lowered detection threshold to obtain up to 2k features per image. Using the built-in Non-Maxima Suppression filter at 3 pixels. Images are resized to 1200 pixels on the largest side. | https://arxiv.org/abs/1712.07629 | https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork | 20-04-23 | is_baseline | |
Submission ID: 00007 L2-Net, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 87.04 Rank: 99/108 |
0.323 Rank: 90/108 |
0.779 Rank: 48/108 |
0.36265 (±0.00090) Rank: 72/108 |
138.97 Rank: 98/108 |
1211.49 Rank: 85/108 |
3.948 Rank: 92/108 |
0.537 Rank: 80/108 |
0.52019 (±0.00120) Rank: 88/108 |
0.44142 Rank: 82/108 |
Challenge organizers (contact) | sift-def | l2net (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-23 | is_baseline | |
Submission ID: 00599 SEKDSize: 512 bytes. Matches: built-in |
2043.44 | 129.46 Rank: 65/108 |
0.386 Rank: 26/108 |
0.809 Rank: 20/108 |
0.45066 (±0.00057) Rank: 26/108 |
176.60 Rank: 71/108 |
1209.63 Rank: 87/108 |
4.437 Rank: 36/108 |
0.454 Rank: 32/108 |
0.66094 (±0.00227) Rank: 33/108 |
0.55580 Rank: 28/108 |
Yafei Song, Ling Cai, Jia Li, Yonghong Tian, Mingyang Li (contact) | sekd | sekd (128 float32: 512 bytes) | Our method named SEKD: Self-Evolving Keypoint Detection and Description, where the SEKD model is trained using COCO validation set. In this submission each image has up to 2048 SEKD keypoints, and 128-dim float descriptor. We use the built-in matcher (bidirectional filter with the both strategy, without flann, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-05-29 | is_submission, is_challenge_2020 | |
Submission ID: 00136 CV-DoG-AffNet-HardNet-kornia-MAG...Size: 512 bytes. Matches: built-in |
2047.84 | 195.17 Rank: 32/108 |
0.339 Rank: 49/108 |
0.774 Rank: 54/108 |
0.41747 (±0.00072) Rank: 47/108 |
284.06 Rank: 38/108 |
1788.70 Rank: 30/108 |
4.191 Rank: 72/108 |
0.511 Rank: 66/108 |
0.58536 (±0.00125) Rank: 64/108 |
0.50141 Rank: 56/108 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00011 GeoDesc, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 93.44 Rank: 96/108 |
0.323 Rank: 90/108 |
0.759 Rank: 70/108 |
0.34215 (±0.00054) Rank: 77/108 |
120.68 Rank: 104/108 |
1098.70 Rank: 97/108 |
4.002 Rank: 89/108 |
0.552 Rank: 86/108 |
0.50594 (±0.00073) Rank: 91/108 |
0.42405 Rank: 84/108 |
Challenge organizers (contact) | sift-def | geodesc (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-23 | is_baseline | |
Submission ID: 00018 R2D2 (r2d2-wasf-n16), DEGENSACSize: 512 bytes. Matches: built-in |
2048.00 | 176.49 Rank: 36/108 |
0.422 Rank: 20/108 |
0.725 Rank: 80/108 |
0.35783 (±0.00061) Rank: 74/108 |
262.02 Rank: 44/108 |
1188.11 Rank: 89/108 |
4.299 Rank: 58/108 |
0.499 Rank: 61/108 |
0.60639 (±0.00124) Rank: 56/108 |
0.48211 Rank: 61/108 |
Challenge organizers (contact) | r2d2-wasf-n16 | r2d2-wasf-n16 (128 float32: 512 bytes) | R2D2 (model r2d2-wasf-n16) with 2k features, using the built-in matcher (bidirectional filter with the both strategy) and no ratio test, and DEGENSAC with optimal parameters for stereo. | N/A | https://opencv.org | 20-04-23 | is_baseline | |
Submission ID: ????? CV-DoG-HardNet8-PTv2Size: 512 bytes. Matches: built-in |
2047.77 | 153.69 Rank: 47/108 |
0.339 Rank: 53/108 |
0.821 Rank: 8/108 |
0.43869 (±0.00017) Rank: 32/108 |
158.63 Rank: 84/108 |
1221.12 Rank: 84/108 |
4.350 Rank: 42/108 |
0.484 Rank: 53/108 |
0.60206 (±0.00003) Rank: 59/108 |
0.52038 Rank: 48/108 |
Milan Pultar, Dmytro Mishkin, Jiri Matas (contact) | sift2k | h8e512pt (128 float32: 512 bytes) | [sid:00619] HardNet8 with PCA compression, batch sampling from few images | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00025 DELF-GLD (32D), DEGENSACSize: 128 bytes. Matches: built-in |
2036.82 | 96.56 Rank: 94/108 |
0.109 Rank: 102/108 |
0.119 Rank: 107/108 |
0.06573 (±0.00030) Rank: 105/108 |
453.19 Rank: 7/108 |
1805.48 Rank: 27/108 |
2.425 Rank: 106/108 |
0.920 Rank: 108/108 |
0.09163 (±0.00140) Rank: 106/108 |
0.07868 Rank: 106/108 |
Challenge organizers (contact) | delf-gld-2k-32d | delf-gld-2k-32d (32 float32: 128 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 32 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with DEGENSAC. | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00578 Key.Net-s + DescNet w/ DEGENSACSize: 512 bytes. Matches: built-in |
2033.67 | 246.62 Rank: 26/108 |
0.449 Rank: 9/108 |
0.805 Rank: 25/108 |
0.45418 (±0.00056) Rank: 24/108 |
331.55 Rank: 24/108 |
1621.68 Rank: 42/108 |
4.570 Rank: 26/108 |
0.447 Rank: 28/108 |
0.67414 (±0.00371) Rank: 28/108 |
0.56416 Rank: 24/108 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet-s | descnet (128 float32: 512 bytes) | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | ||
Submission ID: 00016 LogPolarDesc, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 148.56 Rank: 53/108 |
0.323 Rank: 84/108 |
0.752 Rank: 74/108 |
0.38840 (±0.00064) Rank: 60/108 |
162.84 Rank: 81/108 |
1385.60 Rank: 62/108 |
4.046 Rank: 81/108 |
0.516 Rank: 69/108 |
0.57149 (±0.00344) Rank: 68/108 |
0.47994 Rank: 63/108 |
Challenge organizers (contact) | sift-def | logpolar (128 float32: 512 bytes) | LogPolarDesc descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-04-23 | is_baseline | |
Submission ID: 00704 SOSNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 171.22 Rank: 38/108 |
0.333 Rank: 60/108 |
0.804 Rank: 26/108 |
0.45053 (±0.00053) Rank: 27/108 |
194.04 Rank: 63/108 |
1442.32 Rank: 54/108 |
4.313 Rank: 52/108 |
0.467 Rank: 44/108 |
0.63586 (±0.00246) Rank: 43/108 |
0.54320 Rank: 33/108 |
Challenge organizers (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-06-02 | is_baseline | |
Submission ID: 00604 Key.Net+GIFT+GMCNet+DEGENSACSize: 512 bytes. Matches: custom |
2046.07 | 377.20 Rank: 9/108 |
0.450 Rank: 8/108 |
0.776 Rank: 51/108 |
0.45282 (±0.00000) Rank: 25/108 |
386.63 Rank: 19/108 |
1474.77 Rank: 51/108 |
4.663 Rank: 21/108 |
0.420 Rank: 17/108 |
0.70500 (±0.00294) Rank: 16/108 |
0.57891 Rank: 19/108 |
Yuan Liu (contact) | keynet-2k | scale-gift (128 float32: 512 bytes) | Detecting by Key.Net, descriptors from GIFT, matching by Graph Motion Coherence Network, geometry estimated by DEGENSAC with inlier threshold 0.7 | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00700 HardNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 152.69 Rank: 48/108 |
0.333 Rank: 60/108 |
0.812 Rank: 17/108 |
0.46093 (±0.00080) Rank: 21/108 |
201.26 Rank: 59/108 |
1467.86 Rank: 53/108 |
4.311 Rank: 53/108 |
0.466 Rank: 43/108 |
0.63544 (±0.00324) Rank: 44/108 |
0.54819 Rank: 30/108 |
Challenge organizers (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-06-01 | is_baseline | |
Submission ID: 00036 SOSNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 155.75 Rank: 45/108 |
0.333 Rank: 60/108 |
0.794 Rank: 36/108 |
0.43077 (±0.00017) Rank: 38/108 |
183.92 Rank: 67/108 |
1403.66 Rank: 59/108 |
4.314 Rank: 50/108 |
0.472 Rank: 47/108 |
0.62181 (±0.00161) Rank: 50/108 |
0.52629 Rank: 47/108 |
Challenge organizers (contact) | sift-def | sosnet-upright (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00020 D2-Net (multi-scale), DEGENSACSize: 2048 bytes. Matches: built-in |
2045.55 | 118.41 Rank: 70/108 |
0.168 Rank: 101/108 |
0.293 Rank: 101/108 |
0.12270 (±0.00048) Rank: 98/108 |
286.10 Rank: 36/108 |
1999.37 Rank: 19/108 |
3.013 Rank: 101/108 |
0.774 Rank: 99/108 |
0.28297 (±0.00254) Rank: 99/108 |
0.20283 Rank: 100/108 |
Challenge organizers (contact) | d2net-multiscale | d2net-multiscale (512 float32: 2048 bytes) | D2-Net, multi-scale model, up to 2048 features. Trained on the MegaDepth dataset, removing scenes which overlap with the Phototourism test set. Stereo with DEGENSAC and optimal parameters. Parameters tuned separately from the 8k submission. | http://openaccess.thecvf.com/content_CVPR_2019/papers/Dusmanu_D2-Net_A_Trainable_CNN_for_Joint_Description_and_Detection_of_CVPR_2019_paper.pdf | https://github.com/mihaidusmanu/d2-net | 20-04-23 | is_baseline | |
Submission ID: 00021 DELF-GLD (512D), PyRANSACSize: 2048 bytes. Matches: built-in |
2036.82 | 92.96 Rank: 97/108 |
0.109 Rank: 102/108 |
0.175 Rank: 102/108 |
0.07687 (±0.00010) Rank: 103/108 |
437.64 Rank: 11/108 |
2296.72 Rank: 6/108 |
2.567 Rank: 102/108 |
0.878 Rank: 103/108 |
0.14991 (±0.00513) Rank: 101/108 |
0.11339 Rank: 102/108 |
Challenge organizers (contact) | delf-gld-2k-512d | delf-gld-2k-512d (512 float32: 2048 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 512 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with PyRANSAC (DEGENSAC with the degeneracy check turned off). | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00575 SIFT2k-2048-HardNet64-rain-all-s...Size: 512 bytes. Matches: built-in |
2047.77 | 161.77 Rank: 41/108 |
0.339 Rank: 52/108 |
0.814 Rank: 13/108 |
0.43543 (±0.00051) Rank: 36/108 |
255.45 Rank: 46/108 |
1708.29 Rank: 36/108 |
4.327 Rank: 48/108 |
0.451 Rank: 31/108 |
0.64653 (±0.00221) Rank: 36/108 |
0.54098 Rank: 34/108 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | sift2k-2048-hardnet64-train-all-sos-812000 (128 float32: 512 bytes) | SIFT with 2048 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 812000), FLANN disabled | N/A | N/A | 20-05-20 | is_submission, is_challenge_2020 | |
Submission ID: 00531 Key.Net + X-Net w/ DEGENSACSize: 512 bytes. Matches: built-in |
2040.08 | 173.65 Rank: 37/108 |
0.447 Rank: 12/108 |
0.797 Rank: 33/108 |
0.38608 (±0.00068) Rank: 63/108 |
276.91 Rank: 42/108 |
1468.50 Rank: 52/108 |
4.447 Rank: 35/108 |
0.463 Rank: 39/108 |
0.64369 (±0.00081) Rank: 39/108 |
0.51489 Rank: 50/108 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet | x-net (128 float32: 512 bytes) | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | ||
Submission ID: 00515 SuperPoint + SuperGlue + DEGENSA...Size: 1024 bytes. Matches: custom |
1973.62 | 320.50 Rank: 14/108 |
0.364 Rank: 35/108 |
0.772 Rank: 57/108 |
0.55234 (±0.00024) Rank: 4/108 |
429.48 Rank: 16/108 |
2131.07 Rank: 16/108 |
4.572 Rank: 24/108 |
0.373 Rank: 7/108 |
0.75213 (±0.00345) Rank: 11/108 |
0.65224 Rank: 5/108 |
Paul-Edouard Sarlin (contact) | superpoint-nms3-refine2-r1600forcecubic | superpoint (256 float32: 1024 bytes) | SuperPoint detector and descriptor (2048 keypoints, NMS with radius 3, refinement, on 1600-pixel images). SuperGlue matcher (outdoor model, 150 Sinkhorn iterations). For stereo, DEGENSAC model estimator (1.2 pixels inlier threshold). | https://arxiv.org/abs/1911.11763 | https://github.com/magicleap/SuperGluePretrainedNetwork | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00565 SIFT2k_2000_HardNet64-train-all-...Size: 512 bytes. Matches: built-in |
1999.83 | 157.70 Rank: 43/108 |
0.336 Rank: 54/108 |
0.813 Rank: 16/108 |
0.43188 (±0.00005) Rank: 37/108 |
248.91 Rank: 51/108 |
1666.27 Rank: 38/108 |
4.323 Rank: 49/108 |
0.465 Rank: 41/108 |
0.64404 (±0.00271) Rank: 38/108 |
0.53796 Rank: 38/108 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | sift2k-2000-hardnet64-train-all-sos-812000 (128 float32: 512 bytes) | SIFT with 2000 keypoints(scale 12), sosnet64 with 128 descriptors(trained with sos loss and step 812000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00512 Hardnet with DEGENSAC for stereoSize: 512 bytes. Matches: built-in |
1999.50 | 51.91 Rank: 108/108 |
0.172 Rank: 100/108 |
0.558 Rank: 98/108 |
0.20771 (±0.00049) Rank: 93/108 |
181.48 Rank: 69/108 |
1692.63 Rank: 37/108 |
3.050 Rank: 100/108 |
0.695 Rank: 95/108 |
0.38463 (±0.00275) Rank: 96/108 |
0.29617 Rank: 94/108 |
Vu Trung Nghia & Nguyen Trung Hieu (contact) | siftdef | hardnet-64 (128 float32: 512 bytes) | Using hardnet network to embed a patch (32 x 32) to a 128 (float32) dimensions vector, For stereo, we use DEGENSAC (Chum et al, CVPR'05) with optimal settings | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00012 GeoDesc (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 144.92 Rank: 55/108 |
0.333 Rank: 60/108 |
0.765 Rank: 66/108 |
0.38124 (±0.00056) Rank: 67/108 |
149.66 Rank: 92/108 |
1237.96 Rank: 81/108 |
4.219 Rank: 69/108 |
0.526 Rank: 76/108 |
0.56687 (±0.00261) Rank: 70/108 |
0.47405 Rank: 67/108 |
Challenge organizers (contact) | sift-def | geodesc-upright (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-23 | is_baseline | |
Submission ID: 00038 SOSNet, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 134.59 Rank: 61/108 |
0.323 Rank: 84/108 |
0.760 Rank: 69/108 |
0.38654 (±0.00056) Rank: 62/108 |
145.47 Rank: 96/108 |
1271.41 Rank: 79/108 |
4.045 Rank: 83/108 |
0.526 Rank: 75/108 |
0.56072 (±0.00077) Rank: 75/108 |
0.47363 Rank: 68/108 |
Challenge organizers (contact) | sift-def | sosnet (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00582 R2D2Size: 512 bytes. Matches: built-in |
2048.00 | 155.21 Rank: 46/108 |
0.403 Rank: 23/108 |
0.709 Rank: 82/108 |
0.32501 (±0.00085) Rank: 84/108 |
285.11 Rank: 37/108 |
1162.94 Rank: 93/108 |
4.161 Rank: 75/108 |
0.508 Rank: 63/108 |
0.60276 (±0.00062) Rank: 57/108 |
0.46389 Rank: 77/108 |
(contact) | r2d2-5k-p | r2d2-5k-p (128 float32: 512 bytes) | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | ||
Submission ID: 00555 SuperPoint+GIFT+Graph Motion Coh...Size: 512 bytes. Matches: custom |
1940.85 | 281.06 Rank: 20/108 |
0.356 Rank: 40/108 |
0.676 Rank: 87/108 |
0.42283 (±0.00000) Rank: 43/108 |
290.06 Rank: 34/108 |
1887.70 Rank: 25/108 |
4.567 Rank: 28/108 |
0.420 Rank: 18/108 |
0.70395 (±0.00239) Rank: 18/108 |
0.56339 Rank: 25/108 |
Yuan Liu (contact) | superpoint-2k | scale-gift (128 float32: 512 bytes) | Detecting by SuperPoint, descriptors from GIFT, matching by Graph Motion Coherence Network, geometry estimated by DEGENSAC with inlier threshold 1.0 | N/A | N/A | 20-05-10 | is_submission, is_challenge_2020 | |
Submission ID: 00657 pretrained_sg_DEGENSACSize: 1024 bytes. Matches: custom |
1949.08 | 306.63 Rank: 16/108 |
0.365 Rank: 33/108 |
0.769 Rank: 61/108 |
0.54217 (±0.00000) Rank: 8/108 |
314.15 Rank: 25/108 |
1756.58 Rank: 34/108 |
4.808 Rank: 14/108 |
0.369 Rank: 5/108 |
0.76221 (±0.00290) Rank: 5/108 |
0.65219 Rank: 6/108 |
(contact) | superglue | superglue (256 float32: 1024 bytes) | pretrained outdoor superglue | N/A | N/A | 21-04-24 | is_submission | |
Submission ID: 00640 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: built-in |
2048.00 | 128.76 Rank: 66/108 |
0.450 Rank: 4/108 |
0.744 Rank: 77/108 |
0.33122 (±0.00042) Rank: 82/108 |
289.47 Rank: 35/108 |
1543.70 Rank: 47/108 |
4.284 Rank: 60/108 |
0.509 Rank: 64/108 |
0.60128 (±0.00311) Rank: 60/108 |
0.46625 Rank: 75/108 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features by score. HardNet ,pre-trained on all datasets, is used as a descriptor head. DEGENSA with degeneracy check on is used as a robust estimator. Cyclic consistency matching with a threshold of 0.95 is used. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-02-17 | is_submission | |
Submission ID: 00039 HardNet, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 138.43 Rank: 60/108 |
0.323 Rank: 84/108 |
0.757 Rank: 72/108 |
0.38490 (±0.00120) Rank: 65/108 |
153.29 Rank: 89/108 |
1306.64 Rank: 72/108 |
4.026 Rank: 85/108 |
0.537 Rank: 81/108 |
0.55573 (±0.00175) Rank: 79/108 |
0.47032 Rank: 72/108 |
Challenge organizers (contact) | sift-def | hardnet (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00516 Key.Net (MS 4-1, Pyramid) + X-Ne...Size: 512 bytes. Matches: built-in |
2039.11 | 265.52 Rank: 22/108 |
0.447 Rank: 13/108 |
0.776 Rank: 53/108 |
0.40559 (±0.00112) Rank: 51/108 |
336.81 Rank: 23/108 |
1559.74 Rank: 45/108 |
4.536 Rank: 30/108 |
0.461 Rank: 37/108 |
0.66406 (±0.00521) Rank: 32/108 |
0.53483 Rank: 41/108 |
Barroso-Laguna, Axel and Tian, Yurun and Ng, Tony (contact) | keynet-41-pyramidv2 | x-net-lib (128 float32: 512 bytes) | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | ||
Submission ID: 00517 retrained Hardnet with MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 133.39 Rank: 63/108 |
0.333 Rank: 60/108 |
0.808 Rank: 22/108 |
0.40042 (±0.00077) Rank: 54/108 |
210.10 Rank: 56/108 |
1480.67 Rank: 50/108 |
4.226 Rank: 66/108 |
0.479 Rank: 51/108 |
0.63030 (±0.00242) Rank: 46/108 |
0.51536 Rank: 49/108 |
Vu Trung Nghia and Nguyen Trung Hieu (contact) | sift-upright | hardnet (128 float32: 512 bytes) | Using SIFT upright and retrained Hardnet for keypoints and descriptors. For stereo, we use DEGENSAC Chum et al, CVPR05 with optimal settings | N/A | N/A | 20-04-24 | is_submission, is_challenge_2020 | |
Submission ID: 00711 DISK-32D (depth)Size: 128 bytes. Matches: built-in |
2048.00 | 341.83 Rank: 13/108 |
0.448 Rank: 10/108 |
0.842 Rank: 3/108 |
0.47005 (±0.00037) Rank: 15/108 |
441.16 Rank: 10/108 |
2231.55 Rank: 11/108 |
5.524 Rank: 6/108 |
0.415 Rank: 16/108 |
0.70220 (±0.00208) Rank: 19/108 |
0.58613 Rank: 16/108 |
Michal Tyszkiewicz (contact) | disk-2020-09-15-nms-7-depth-32-save-46-imsize-1024-nump-2048 | disk-2020-09-15-nms-7-depth-32-save-46-imsize-1024-nump-2048 (32 float32: 128 bytes) | Local feature model learned via policy gradient, using 32D descriptors. Model trained with a cycle-consistency loss and supervised with depth-based constraints. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 7x7 window, taking the top 2048 features by score. | N/A | N/A | 20-09-16 | is_baseline | |
Submission ID: 00009 GeoDesc (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 116.56 Rank: 75/108 |
0.333 Rank: 60/108 |
0.784 Rank: 42/108 |
0.38660 (±0.00078) Rank: 61/108 |
149.66 Rank: 92/108 |
1237.96 Rank: 81/108 |
4.219 Rank: 69/108 |
0.525 Rank: 74/108 |
0.56687 (±0.00261) Rank: 70/108 |
0.47673 Rank: 65/108 |
Challenge organizers (contact) | sift-def | geodesc-upright (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-22 | is_baseline | |
Submission ID: 00623 HardNet-epoch2-OANetSize: 512 bytes. Matches: custom |
2047.76 | 247.45 Rank: 25/108 |
0.347 Rank: 43/108 |
0.765 Rank: 65/108 |
0.44717 (±0.00000) Rank: 29/108 |
254.11 Rank: 48/108 |
1898.45 Rank: 24/108 |
4.503 Rank: 32/108 |
0.437 Rank: 26/108 |
0.67958 (±0.00183) Rank: 26/108 |
0.56337 Rank: 26/108 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift2k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 2k features, using the oanet trained from scratch and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-06-02 | is_submission, is_challenge_2020 | |
Submission ID: 00670 superglue_adapt_degensacSize: 1024 bytes. Matches: custom |
2048.00 | 365.61 Rank: 10/108 |
0.383 Rank: 28/108 |
0.747 Rank: 76/108 |
0.50988 (±0.00000) Rank: 13/108 |
376.42 Rank: 20/108 |
2076.46 Rank: 18/108 |
4.905 Rank: 13/108 |
0.373 Rank: 6/108 |
0.76226 (±0.00085) Rank: 4/108 |
0.63607 Rank: 11/108 |
Downey (contact) | superpointadapt30 | superglueadapt30 (256 float32: 1024 bytes) | degensac t1.2 adaption30 | N/A | N/A | 21-05-07 | is_submission | |
Submission ID: 00620 pffNet + SuperPoint + MAGSACSize: 512 bytes. Matches: built-in |
1267.22 | 57.48 Rank: 106/108 |
0.341 Rank: 47/108 |
0.632 Rank: 90/108 |
0.22549 (±0.00077) Rank: 92/108 |
131.91 Rank: 100/108 |
893.88 Rank: 104/108 |
4.339 Rank: 43/108 |
0.531 Rank: 78/108 |
0.55935 (±0.00143) Rank: 77/108 |
0.39242 Rank: 91/108 |
Jongmin Lee, Seungwook Kim, Yoonwoo Jeong (contact) | superpoint | pffnet (128 float32: 512 bytes) | pffNet descriptors + SuperPoint keypoints + MAGSAC outlier-filtering | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | |
Submission ID: 00135 CV-DoG-AffNet-HardNet-kornia-PyR...Size: 512 bytes. Matches: built-in |
2047.84 | 105.57 Rank: 87/108 |
0.339 Rank: 49/108 |
0.800 Rank: 30/108 |
0.35887 (±0.00090) Rank: 73/108 |
284.06 Rank: 38/108 |
1788.70 Rank: 30/108 |
4.191 Rank: 72/108 |
0.513 Rank: 67/108 |
0.58536 (±0.00125) Rank: 64/108 |
0.47211 Rank: 69/108 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00026 AKAZE (OpenCV), DEGENSACSize: 61 bytes. Matches: built-in |
1657.83 | 69.77 Rank: 104/108 |
0.355 Rank: 41/108 |
0.622 Rank: 95/108 |
0.17632 (±0.00087) Rank: 96/108 |
224.40 Rank: 54/108 |
795.54 Rank: 107/108 |
3.311 Rank: 99/108 |
0.819 Rank: 100/108 |
0.23020 (±0.00466) Rank: 100/108 |
0.20326 Rank: 99/108 |
Challenge organizers (contact) | akaze-def | akaze (61 uint8: 61 bytes) | AKAZE with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00706 LogPolar-Upright w/ DEGENSAC (no...Size: 512 bytes. Matches: built-in |
1892.71 | 162.20 Rank: 40/108 |
0.333 Rank: 60/108 |
0.807 Rank: 23/108 |
0.45674 (±0.00059) Rank: 23/108 |
211.87 Rank: 55/108 |
1553.36 Rank: 46/108 |
4.331 Rank: 46/108 |
0.471 Rank: 46/108 |
0.63701 (±0.00278) Rank: 42/108 |
0.54688 Rank: 31/108 |
Challenge organizers (contact) | sift-def | logpolar-upright (128 float32: 512 bytes) | Upright LogPolar descriptors on DoG features (OpenCV). Using the built-in matcher: bidirectional filter with the both strategy. DEGENSAC for stereo. FLANN disabled. | https://arxiv.org/abs/1908.05547 | https://github.com/cvlab-epfl/log-polar-descriptors | 20-06-02 | is_baseline | |
Submission ID: 00669 disk_degree(refine)_End-to-EndSize: 512 bytes. Matches: custom |
2048.00 | 591.03 Rank: 1/108 |
0.447 Rank: 15/108 |
0.811 Rank: 18/108 |
0.53291 (±0.00000) Rank: 10/108 |
602.70 Rank: 1/108 |
2633.91 Rank: 1/108 |
5.556 Rank: 4/108 |
0.382 Rank: 10/108 |
0.75413 (±0.00231) Rank: 9/108 |
0.64352 Rank: 10/108 |
Weiyue Zhao (contact) | disk | disk (128 float32: 512 bytes) | disk discriptors, followed by degree(refine)_End-to-End and DEGENSAC. | N/A | N/A | 21-05-07 | is_submission | |
Submission ID: 00662 retrained_superglue_DEGENSACSize: 1024 bytes. Matches: custom |
1949.08 | 306.63 Rank: 16/108 |
0.365 Rank: 33/108 |
0.769 Rank: 61/108 |
0.54220 (±0.00000) Rank: 7/108 |
314.15 Rank: 25/108 |
1757.90 Rank: 33/108 |
4.805 Rank: 15/108 |
0.375 Rank: 9/108 |
0.76055 (±0.00173) Rank: 6/108 |
0.65137 Rank: 8/108 |
(contact) | superglue | superglue (256 float32: 1024 bytes) | pretrained outdoor superglue | N/A | N/A | 21-04-25 | is_submission | |
Submission ID: 00022 L2-Net (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 110.72 Rank: 82/108 |
0.333 Rank: 60/108 |
0.801 Rank: 29/108 |
0.40729 (±0.00097) Rank: 50/108 |
171.93 Rank: 73/108 |
1333.56 Rank: 68/108 |
4.226 Rank: 67/108 |
0.497 Rank: 58/108 |
0.58762 (±0.00228) Rank: 62/108 |
0.49746 Rank: 58/108 |
Challenge organizers (contact) | sift-def | l2net-upright (128 float32: 512 bytes) | L2-Net descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf | https://github.com/vcg-uvic/image-matching-benchmark-baselines | 20-04-23 | is_baseline | |
Submission ID: 00606 HardNet-epoch2-OANetSize: 512 bytes. Matches: custom |
2047.76 | 244.62 Rank: 28/108 |
0.347 Rank: 43/108 |
0.773 Rank: 56/108 |
0.46389 (±0.00000) Rank: 20/108 |
251.11 Rank: 50/108 |
1916.44 Rank: 23/108 |
4.501 Rank: 33/108 |
0.428 Rank: 23/108 |
0.69407 (±0.00216) Rank: 23/108 |
0.57898 Rank: 18/108 |
Chen Shen, Zhipeng Wang, Jun Zhang, Zhongkun Chen, Zhiwei Ruan, Jingchao Zhou, Pengfei Xu (contact) | sift2k | hardnet-epoch2 (128 float32: 512 bytes) | sift and hardnet with 2k features, using the oanet and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-05-30 | is_submission, is_challenge_2020 | |
Submission ID: 00013 SIFT (OpenCV), DEGENSACSize: 512 bytes. Matches: built-in |
1936.30 | 76.94 Rank: 101/108 |
0.323 Rank: 94/108 |
0.733 Rank: 79/108 |
0.30766 (±0.00075) Rank: 86/108 |
171.92 Rank: 75/108 |
1059.70 Rank: 101/108 |
3.687 Rank: 96/108 |
0.618 Rank: 94/108 |
0.42476 (±0.00103) Rank: 94/108 |
0.36621 Rank: 92/108 |
Challenge organizers (contact) | sift-def | sift (128 float32: 512 bytes) | SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-23 | is_baseline | |
Submission ID: 00513 Personal Inplementation SIFT (GP...Size: 512 bytes. Matches: built-in |
1900.60 | 70.04 Rank: 103/108 |
0.085 Rank: 108/108 |
0.002 Rank: 108/108 |
0.07782 (±0.00022) Rank: 102/108 |
173.06 Rank: 72/108 |
948.78 Rank: 103/108 |
3.816 Rank: 94/108 |
0.887 Rank: 106/108 |
0.09835 (±0.00068) Rank: 105/108 |
0.08809 Rank: 105/108 |
feyman_priv (contact) | sift-gpu | sift-gpu (128 float32: 512 bytes) | SIFT with 2048 features, using the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC, and setting keypoint orientation to a constant value to increase performance. | N/A | N/A | 20-04-25 | is_submission, is_challenge_2020 | |
Submission ID: 00609 Key.Net-s-ref + HyNet w/ DEGENSA...Size: 512 bytes. Matches: built-in |
2032.93 | 270.04 Rank: 21/108 |
0.461 Rank: 2/108 |
0.820 Rank: 10/108 |
0.46902 (±0.00032) Rank: 16/108 |
280.40 Rank: 41/108 |
1489.62 Rank: 49/108 |
4.689 Rank: 20/108 |
0.436 Rank: 25/108 |
0.68120 (±0.00119) Rank: 25/108 |
0.57511 Rank: 21/108 |
Barroso-Laguna, Axel and Tian, Yurun, Ng, Tony and Mikolajczyk, Krystian (contact) | keynet-s-ref | hynet (128 float32: 512 bytes) | N/A | N/A | 20-06-01 | is_submission, is_challenge_2020 | ||
Submission ID: 00033 SOSNet, DEGENSACSize: 512 bytes. Matches: built-in |
1936.28 | 119.91 Rank: 69/108 |
0.323 Rank: 84/108 |
0.769 Rank: 63/108 |
0.37632 (±0.00056) Rank: 68/108 |
145.47 Rank: 96/108 |
1271.41 Rank: 79/108 |
4.045 Rank: 83/108 |
0.524 Rank: 73/108 |
0.56072 (±0.00077) Rank: 75/108 |
0.46852 Rank: 74/108 |
Challenge organizers (contact) | sift-def | sosnet (128 float32: 512 bytes) | SOSNet descriptors extracted on SIFT keypoints and DEGENSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1904.05019 | https://github.com/yuruntian/SOSNet | 20-04-22 | is_baseline | |
Submission ID: 00701 GeoDesc (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 132.69 Rank: 64/108 |
0.333 Rank: 60/108 |
0.798 Rank: 32/108 |
0.41364 (±0.00033) Rank: 49/108 |
161.05 Rank: 83/108 |
1287.26 Rank: 76/108 |
4.239 Rank: 64/108 |
0.502 Rank: 62/108 |
0.58373 (±0.00402) Rank: 67/108 |
0.49868 Rank: 57/108 |
Challenge organizers (contact) | sift-def | geodesc-upright (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-06-01 | is_baseline | |
Submission ID: 00040 DELF-GLD (128D), DEGENSACSize: 512 bytes. Matches: built-in |
2036.82 | 100.78 Rank: 91/108 |
0.109 Rank: 102/108 |
0.151 Rank: 105/108 |
0.08367 (±0.00047) Rank: 100/108 |
430.68 Rank: 13/108 |
2162.72 Rank: 13/108 |
2.535 Rank: 104/108 |
0.880 Rank: 104/108 |
0.13247 (±0.00116) Rank: 103/108 |
0.10807 Rank: 103/108 |
Challenge organizers (contact) | delf-gld-2k-128d | delf-gld-2k-128d (128 float32: 512 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 128 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with DEGENSAC. | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00564 SIFT2k_2000_HardNet64_train_all_...Size: 512 bytes. Matches: built-in |
1999.83 | 156.93 Rank: 44/108 |
0.336 Rank: 54/108 |
0.814 Rank: 14/108 |
0.42959 (±0.00060) Rank: 39/108 |
244.13 Rank: 52/108 |
1646.76 Rank: 41/108 |
4.334 Rank: 45/108 |
0.459 Rank: 36/108 |
0.64407 (±0.00339) Rank: 37/108 |
0.53683 Rank: 39/108 |
Ximin Zheng, Sheng He, Hualong Shi (contact) | sift2k | sift2k-2000-hardnet64-train-all-l2-138000 (128 float32: 512 bytes) | SIFT with 2000 keypoints(scale 12), hardnet64 with 128 descriptors(trained with l2 loss and step 138000), FLANN disabled | N/A | N/A | 20-05-13 | is_submission, is_challenge_2020 | |
Submission ID: 00616 Sift-Fusion_Avg-NM-Net_End-to-En...Size: 512 bytes. Matches: custom |
1892.70 | 142.38 Rank: 56/108 |
0.333 Rank: 76/108 |
0.801 Rank: 28/108 |
0.44818 (±0.00000) Rank: 28/108 |
146.58 Rank: 94/108 |
1339.45 Rank: 67/108 |
4.406 Rank: 40/108 |
0.456 Rank: 33/108 |
0.65563 (±0.00209) Rank: 34/108 |
0.55190 Rank: 29/108 |
Chen Zhao (contact) | siftdef | fusion-avg (128 float32: 512 bytes) | SIFT and Fusion_Avg, followed by NM-Net_End-to-End and DEGENSAC. | N/A | N/A | 20-05-31 | is_submission, is_challenge_2020 | |
Submission ID: 00003 HardNet (upright), DEGENSACSize: 512 bytes. Matches: built-in |
1892.71 | 141.66 Rank: 57/108 |
0.333 Rank: 60/108 |
0.803 Rank: 27/108 |
0.44235 (±0.00127) Rank: 31/108 |
193.09 Rank: 64/108 |
1435.93 Rank: 56/108 |
4.302 Rank: 56/108 |
0.474 Rank: 49/108 |
0.62275 (±0.00423) Rank: 48/108 |
0.53255 Rank: 42/108 |
Challenge organizers (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints with a fixed orientation, and DEGENSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-22 | is_baseline | |
Submission ID: 00030 Upright Root-SIFT (OpenCV), DEGE...Size: 512 bytes. Matches: built-in |
1892.72 | 111.69 Rank: 81/108 |
0.333 Rank: 56/108 |
0.781 Rank: 47/108 |
0.39547 (±0.00027) Rank: 57/108 |
201.15 Rank: 60/108 |
1347.75 Rank: 65/108 |
4.091 Rank: 77/108 |
0.522 Rank: 72/108 |
0.56078 (±0.00264) Rank: 74/108 |
0.47812 Rank: 64/108 |
Challenge organizers (contact) | sift-def | rootsift-upright (128 float32: 512 bytes) | Upright Root-SIFT with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC, and setting keypoint orientation to a constant value. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00027 DELF-GLD (32D), PyRANSACSize: 128 bytes. Matches: built-in |
2036.82 | 91.18 Rank: 98/108 |
0.109 Rank: 102/108 |
0.122 Rank: 106/108 |
0.06197 (±0.00032) Rank: 107/108 |
453.19 Rank: 7/108 |
1805.48 Rank: 27/108 |
2.425 Rank: 106/108 |
0.916 Rank: 107/108 |
0.09163 (±0.00140) Rank: 106/108 |
0.07680 Rank: 107/108 |
Challenge organizers (contact) | delf-gld-2k-32d | delf-gld-2k-32d (32 float32: 128 bytes) | DELF-GLD, with up to 2k features. Descriptors are cropped to 32 dimensions with PCA. Re-using optimal parameters for the (default) 40D models. Stereo with PyRANSAC (DEGENSAC with the degeneracy check turned off). | https://arxiv.org/abs/1812.01584 | https://github.com/tensorflow/models/tree/master/research/delf | 20-04-22 | is_baseline | |
Submission ID: 00037 GeoDesc, MAGSACSize: 512 bytes. Matches: built-in |
1936.28 | 115.17 Rank: 78/108 |
0.323 Rank: 90/108 |
0.742 Rank: 78/108 |
0.34024 (±0.00050) Rank: 79/108 |
120.68 Rank: 104/108 |
1098.70 Rank: 97/108 |
4.002 Rank: 89/108 |
0.551 Rank: 85/108 |
0.50594 (±0.00073) Rank: 91/108 |
0.42309 Rank: 85/108 |
Challenge organizers (contact) | sift-def | geodesc (128 float32: 512 bytes) | GeoDesc descriptors extracted on SIFT keypoints and MAGSAC for stereo. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1807.06294 | https://github.com/lzx551402/geodesc | 20-04-23 | is_baseline | |
Submission ID: 00586 KeyNet+HardNet+NM-Net_v2Size: 512 bytes. Matches: custom |
2042.89 | 160.25 Rank: 42/108 |
0.435 Rank: 18/108 |
0.821 Rank: 7/108 |
0.40264 (±0.00000) Rank: 52/108 |
165.01 Rank: 80/108 |
1134.69 Rank: 95/108 |
4.720 Rank: 19/108 |
0.448 Rank: 30/108 |
0.67488 (±0.00197) Rank: 27/108 |
0.53876 Rank: 35/108 |
Chen Zhao (contact) | keynet | hardnet (128 float32: 512 bytes) | KeyNet and HardNet, followed by NM-Net_v2 and DEGENSAC. | N/A | N/A | 20-05-23 | is_submission, is_challenge_2020 | |
Submission ID: 00046 D2-Net (single-scale), DEGENSACSize: 2048 bytes. Matches: built-in |
2038.22 | 147.53 Rank: 54/108 |
0.202 Rank: 99/108 |
0.337 Rank: 100/108 |
0.16324 (±0.00082) Rank: 97/108 |
364.68 Rank: 21/108 |
1984.98 Rank: 21/108 |
3.412 Rank: 98/108 |
0.720 Rank: 98/108 |
0.36911 (±0.00167) Rank: 98/108 |
0.26618 Rank: 97/108 |
Challenge organizers (contact) | d2net-singlescale | d2net-singlescale (512 float32: 2048 bytes) | D2-Net, single-scale model, up to 2048 features. Trained on the MegaDepth dataset, removing scenes which overlap with the Phototourism test set. Stereo with DEGENSAC and optimal parameters. Parameters tuned separately from the 8k submission. | http://openaccess.thecvf.com/content_CVPR_2019/papers/Dusmanu_D2-Net_A_Trainable_CNN_for_Joint_Description_and_Detection_of_CVPR_2019_paper.pdf | https://github.com/mihaidusmanu/d2-net | 20-04-23 | is_baseline | |
Submission ID: 00710 DISK-32D (epi)Size: 128 bytes. Matches: built-in |
2048.00 | 352.43 Rank: 11/108 |
0.459 Rank: 3/108 |
0.826 Rank: 5/108 |
0.45946 (±0.00023) Rank: 22/108 |
484.30 Rank: 4/108 |
2324.68 Rank: 5/108 |
5.589 Rank: 3/108 |
0.427 Rank: 21/108 |
0.69938 (±0.00192) Rank: 21/108 |
0.57942 Rank: 17/108 |
Michal Tyszkiewicz (contact) | disk-2020-09-15-nms-5-epi-32-save-49-imsize-1024-nump-2048 | disk-2020-09-15-nms-5-epi-32-save-49-imsize-1024-nump-2048 (32 float32: 128 bytes) | Local feature model learned via policy gradient, using 32D descriptors. Model trained with a cycle-consistency loss and supervised with epipolar constraints. Trained on MegaDepth, removing conflicts with the test data. For inference, images are resized to 1024 pixels on the longest edge, with NMS over a 5x5 window, taking the top 2048 features by score. | N/A | N/A | 20-09-16 | is_baseline | |
Submission ID: 00642 MT-2-Hardnet-Pretraind-all-Datas...Size: 512 bytes. Matches: custom |
2048.00 | 140.36 Rank: 58/108 |
0.450 Rank: 4/108 |
0.571 Rank: 97/108 |
0.19708 (±0.00000) Rank: 95/108 |
146.55 Rank: 95/108 |
1279.68 Rank: 77/108 |
4.057 Rank: 79/108 |
0.542 Rank: 84/108 |
0.53330 (±0.00263) Rank: 83/108 |
0.36519 Rank: 93/108 |
Anonymous (to be released: 2020-6-12) | mt-2 | hardnet (128 float32: 512 bytes) | Local feature model learned via training with covariant constraint loss function. We take the top 2048 features by score. HardNet ,pre-trained on all datasets, is used as a descriptor head. Ordered-Aware(OA) Network, trained from scratch on sift-side-8k features (inlier threshold of network output = 1.0, post process method = None, with other settings set to default), is used to compute robust matches and OANet is used in place of robust model estimator. | Anonymous (to be released: 2020-6-12) | Anonymous (to be released: 2020-6-12) | 21-03-07 | is_submission | |
Submission ID: 00134 CV-DoG-AffNet-HardNet-kornia-DEG...Size: 512 bytes. Matches: built-in |
2047.84 | 152.12 Rank: 50/108 |
0.339 Rank: 49/108 |
0.791 Rank: 37/108 |
0.41972 (±0.00027) Rank: 45/108 |
284.06 Rank: 38/108 |
1788.70 Rank: 30/108 |
4.191 Rank: 72/108 |
0.511 Rank: 65/108 |
0.58536 (±0.00125) Rank: 64/108 |
0.50254 Rank: 55/108 |
Challenge organizers (contact) | sift8k | affnethardnet (128 float32: 512 bytes) | OpenCV DoG keypoints, followed by AffNet shape estimation and HardNet descriptor. Implementation: OpenCV + kornia library | https://arxiv.org/abs/1711.06704 | https://kornia.readthedocs.io/en/latest/feature.html | 21-02-05 | is_baseline | |
Submission ID: 00029 SuperPoint (2k features, NMS=4),...Size: 1024 bytes. Matches: built-in |
2048.00 | 122.69 Rank: 68/108 |
0.364 Rank: 37/108 |
0.630 Rank: 91/108 |
0.28965 (±0.00063) Rank: 87/108 |
170.60 Rank: 76/108 |
1185.38 Rank: 90/108 |
4.331 Rank: 47/108 |
0.557 Rank: 87/108 |
0.54663 (±0.00218) Rank: 81/108 |
0.41814 Rank: 86/108 |
Challenge organizers (contact) | superpoint-nms4-r1200 | superpoint-nms4-r1200 (256 float32: 1024 bytes) | SuperPoint (initial release based on Pytorch). Lowered detection threshold to obtain up to 2k features per image. Using the built-in Non-Maxima Suppression filter at 2 pixels. Images are resized to 1200 pixels on the largest side. | https://arxiv.org/abs/1712.07629 | https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork | 20-04-22 | is_baseline | |
Submission ID: 00017 ORB (OpenCV), DEGENSACSize: 32 bytes. Matches: built-in |
2031.83 | 63.43 Rank: 105/108 |
0.366 Rank: 32/108 |
0.523 Rank: 99/108 |
0.08243 (±0.00058) Rank: 101/108 |
120.93 Rank: 103/108 |
280.38 Rank: 108/108 |
2.191 Rank: 108/108 |
0.863 Rank: 101/108 |
0.02318 (±0.00229) Rank: 108/108 |
0.05280 Rank: 108/108 |
Challenge organizers (contact) | orb | orb (32 uint8: 32 bytes) | ORB with (up to) 2048 features, using the built-in matcher (bidirectional filter with the both strategy) and DEGENSAC. | N/A | https://opencv.org | 20-04-22 | is_baseline | |
Submission ID: 00523 SEKDSize: 512 bytes. Matches: built-in |
2043.44 | 116.81 Rank: 74/108 |
0.386 Rank: 26/108 |
0.799 Rank: 31/108 |
0.42811 (±0.00096) Rank: 40/108 |
166.41 Rank: 79/108 |
1163.31 Rank: 92/108 |
4.427 Rank: 37/108 |
0.456 Rank: 34/108 |
0.64919 (±0.00171) Rank: 35/108 |
0.53865 Rank: 36/108 |
Yafei Song, Ling Cai, Mingyang Li (contact) | sekd | sekd (128 float32: 512 bytes) | Our method named SEKD: Self-Evolving Keypoint Detection and Description, where the SEKD model is trained using COCO test images. In this submission each image has up to 2048 SEKD keypoints, and 128-dim float descriptor. We use the built-in matcher (bidirectional filter with the both strategy, optimal inlier and ratio test thresholds) with DEGENSAC. | N/A | N/A | 20-04-29 | is_submission, is_challenge_2020 | |
Submission ID: 00651 sp_ae_sg_degensacSize: 512 bytes. Matches: custom |
1873.96 | 288.87 Rank: 19/108 |
0.323 Rank: 81/108 |
0.693 Rank: 85/108 |
0.44506 (±0.00000) Rank: 30/108 |
295.52 Rank: 33/108 |
1994.42 Rank: 20/108 |
4.730 Rank: 17/108 |
0.395 Rank: 13/108 |
0.73681 (±0.00306) Rank: 13/108 |
0.59093 Rank: 15/108 |
(contact) | superpoint | superpoint-down128 (128 float32: 512 bytes) | SP with 2048 features, and down load. | N/A | N/A | 21-04-05 | is_submission | |
Submission ID: 00035 HardNet (upright), MAGSACSize: 512 bytes. Matches: built-in |
1892.71 | 181.81 Rank: 34/108 |
0.333 Rank: 60/108 |
0.781 Rank: 46/108 |
0.43816 (±0.00058) Rank: 34/108 |
193.09 Rank: 64/108 |
1435.93 Rank: 56/108 |
4.302 Rank: 56/108 |
0.473 Rank: 48/108 |
0.62275 (±0.00423) Rank: 48/108 |
0.53045 Rank: 44/108 |
Challenge organizers (contact) | sift-def | hardnet-upright (128 float32: 512 bytes) | HardNet descriptors extracted on SIFT keypoints with a fixed orientation, and MAGSAC for stereo. Re-using parameters found for the orientation-sensitive variant. Please refer to the baselines repository (linked) for details. | https://arxiv.org/abs/1705.10872 | https://github.com/DagnyT/hardnet | 20-04-23 | is_baseline |