Download links (GoogleUrban datset is password protected, please contact organizer at for the password):

The basic principle behind our approach is to obtain dense and accurate 3D reconstructions from large collections of images with off-the-shelf Structure from Motion (SfM), such as Colmap or RealityCapture. This provides us with (pseudo-) ground truth poses and, optionally, densified depth maps. We then subsample these image collections into smaller subsets (e.g. pairs of images for stereo, or 5 to 25 images for multiview reconstruction with SfM) and evaluate different methods against the "ground-truth". We publish up to 100 images per scene, after subsampling.

This year we have 3 datasets: "Phototourism", "PragueParks" and "GoogleUrban". Each of them is obtained in a different way, which we will describe below. The "Phototourism" dataset is unchanged from the 2020 challenge, and comes with a training set, along with validation data with a public ground truth, and test data with a private ground truth. PragueParks and GoogleUrban provide validation data, but not training data. Our experiments show that the optimal hyperparameters (such as the RANSAC threshold or the ratio test) vary among datasets, which is why we encourage participants to tune their methods on the each of 3 validation sets, separately, before submitting.

The PragueParks dataset

The PragueParks dataset contains images from video sequences captured by the organizers with an iPhone 11, in 2021. The iPhone 11 has two cameras, with normal and wide lenses, both of which were used. Note that while the video is high quality, some of the frames suffer from motion blur. These videos were then processed by the commercial 3D reconstruction software RealityCapture, which is orders of magnitude faster than COLMAP, while delivering a comparable output in terms of accuracy. Similarly to we did for the "PhotoTourism" dataset, this data is then subsampled in order to generate the subsets used for evaluation.

The dataset contains small-scale scenes like tree, pond, wooden and metal sculptures with different level of zoom, lots of vegetation, and no people. The distribution of its camera poses differs from Phototourism. Learned detectors like SuperPoint and R2D2 work better on this subset than Difference-of-Gaussians used in a SIFT: baselines will be released soon.

Prague Parks Dataset
Validation scene Num. images Num. 3D points
Wooden Lady 2049 1484682
Total 2k 1.5M

Test scenes Num. images Num. 3D points
Tree 614 578252
Pond 3654 1435656
Lizard 264 419477
Total 4532 2.4M

The GoogleUrban dataset

The GoogleUrban dataset contains images used by Google to evaluate localization algorithms, such as those in Google's Visual Positioning System, which powers Live View on millions on mobile devices. They are obtained from videos collected from different cell phones, on many countries all over the world, often years apart. They contain poses, but not depth maps. Please note that due to legal reasons, this data is released with a restricted license, and must be deleted by the end of the challenge.

GoogleUrban Dataset
Validation scenes Num. images
Edinburgh 75
Mexico DF 75
Vancouver 74

Test scenes Num. images
Amsterdam 64
Bangkok 75
Barcelona 75
Buenos Aires 64
Cambridge (MA) 75
Cannes 75
Chicago 75
Helsinki 75
Madrid 75
Mountain View 75
New Orleans 75
San Francisco 75
Singapore 75
Sydney 75
Tokyo 75
Toronto 75
Zurich 75

The Phototourism dataset

This was the dataset used in the previous two versions of the challenge, which remains one of our three datasets. We publish training data with images, poses, depth maps, and co-visibility estimates. We also provide a validation set in the format expected by the benchmark, to allow challenge participants to tune their methods before submission. The test set will remain private.

Reconstruction Images

In order to learn and evaluate models that can perform well under a wide range of situations, it is of paramount importance to collect information from multiple sensors obtained at different times, from different viewpoints, and with occlusions. A natural solution is thus to turn to photo-tourism data. In this dataset we rely on 26 photo-tourism image collections of popular landmarks originally collected by the Yahoo Flickr Creative Commons 100M (YFCC) dataset and Reconstructing the world in six days. They range from ~100s to ~1000s of images.

We provide the full data, with ground truth, for 15 scenes, which can be optionally used for training. We format 3 of them (Reichstag, Sacre Coeur and Saint Peter's Square) for validation, and reserve 12 for testing.

We provide examples to parse the training data on the benchmark repository: please refer to this notebook for details.

Training scene Num. images Num. 3D points
Brandenburg Gate 1363 100040
Buckingham Palace 1676 234052
Colosseum Exterior 2063 259807
Grand Place Brussels 1083 229788
Hagia Sophia Interior 888 235541
Notre Dame Front Facade 3765 488895
Palace of Westminster 983 115868
Pantheon Exterior 1401 166923
Reichstag 75 17823
Sacre Coeur 1179 140659
Saint Peter's Square 2504 232329
Taj Mahal 1312 94121
Temple Nara Japan 904 92131
Trevi Fountain 3191 580673
Westminster Abbey 1061 198222
Total 25.6k 3.7M

Test scenes Num. images Num. 3D points
British Museum 660 73569
Florence Cathedral Side 108 44143
Lincoln Memorial Statue 850 58661
London Bridge 629 72235
Milan Cathedral 124 33905
Mount Rushmore 138 45350
Piazza San Marco 249 95895
Sagrada Familia 401 120723
Saint Paul's Cathedral 615 98872
Total 4107 696k