SuperPoint-128d-adapt + SuperGlue + DEGENSAC

This page ranks the submission against all others using the same number of keypoints, regardless of descriptor size. Please hover over table headers for descriptions on metrics and full scene names.

• Phototourism: Stereo track
• Phototourism: Multiview track

Metadata

• Authors: Paul-Edouard Sarlin (contact)
• Keypoint: superpoint-k2048-nms4-refine2-r1600forcecubic-masked-d.001-adapt50
• Descriptor: superpoint-down128 (128 float32: 512 bytes)
• Number of features: 2048
• Summary: 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).
• Paper: https://arxiv.org/abs/1911.11763
• Website: https://psarlin.com/superglue
• Origin: Submission
• Flags: is_submission, is_challenge_2020

Phototourism / Stereo track

{
    "use_custom_matches": true,
    "custom_matches_name": "superglue-t.2-it150-degensac-th1.1",
    "geom": {
        "method": "cv2-8pt"
    }
}

We show the inliers that survive the robust estimation loop (i.e. RANSAC), or those supplied with the submission if using custom matches, and use the depth estimates to determine whether they are correct. We draw matches above a 5-pixel error threshold in red, and those below are color-coded by their error, from 0 (green) to 5 pixels (yellow). Matches for which we do not have depth estimates are drawn in blue. Please note that the depth maps are estimates and may contain errors.

---

— british museum —

— florence cathedral side —

— lincoln memorial statue —

— london bridge —

— milan cathedral —

— mount rushmore —

— piazza san marco —

— sagrada familia —

— saint paul's cathedral —

Phototourism / Multiview track

{
    "use_custom_matches": true,
    "custom_matches_name": "superglue-t.2-it150-degensac-th1.1",
    "colmap": {}
}

In the multi-view track we reconstruct the scene with Structure-from-Motion (Colmap) with small sets of images. We show the results for one bag of 25 images (displaying: 10). Keypoints are drawn in blue if they are part of the model, and in red otherwise.

---

— british museum —

— florence cathedral side —

— lincoln memorial statue —

— london bridge —

— milan cathedral —

— mount rushmore —

— piazza san marco —

— sagrada familia —

— saint paul's cathedral —