IMW 2020: Leaderboard

Current version: 3623b2ac (2020-02-20, 16:09 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)


Prize #2: restricted keypoints (2k) / standard descriptors (512 bytes)



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)
Authors Keypoint Descriptor Summary Paper Website Processing date
SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7874.85 226.69
Rank: 5/7
0.470
Rank: 5/7
0.809
Rank: 4/7
0.44473
(±0.00120)
Rank: 4/7
399.54
Rank: 7/7
3265.87
Rank: 4/7
3.883
Rank: 4/7
0.561
Rank: 4/7
0.54747
(±0.00225)
Rank: 4/7
0.49610
Rank: 4/7
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 20-01-28
Upright Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7846.57 336.38
Rank: 1/7
0.485
Rank: 3/7
0.830
Rank: 2/7
0.50131
(±0.00017)
Rank: 1/7
516.33
Rank: 3/7
4113.38
Rank: 1/7
4.260
Rank: 1/7
0.490
Rank: 1/7
0.63342
(±0.00110)
Rank: 1/7
0.56737
Rank: 1/7
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 20-01-28
Upright SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7846.57 302.54
Rank: 2/7
0.485
Rank: 3/7
0.815
Rank: 3/7
0.47840
(±0.00047)
Rank: 3/7
499.07
Rank: 4/7
3865.68
Rank: 2/7
4.139
Rank: 2/7
0.524
Rank: 3/7
0.60037
(±0.00279)
Rank: 2/7
0.53938
Rank: 2/7
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 20-01-28
AKAZE (OpenCV), DEGENSAC
Size: 61 bytes. Matches: built-in
7871.55 236.04
Rank: 4/7
0.552
Rank: 1/7
0.723
Rank: 5/7
0.30219
(±0.00109)
Rank: 5/7
471.17
Rank: 5/7
2585.81
Rank: 6/7
3.327
Rank: 6/7
0.772
Rank: 5/7
0.32662
(±0.00278)
Rank: 6/7
0.31441
Rank: 5/7
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 20-01-28
SURF (OpenCV), DEGENSAC
Size: 256 bytes. Matches: built-in
7750.71 121.14
Rank: 7/7
0.433
Rank: 7/7
0.667
Rank: 6/7
0.24110
(±0.00061)
Rank: 6/7
913.30
Rank: 1/7
3217.44
Rank: 5/7
3.353
Rank: 5/7
0.780
Rank: 6/7
0.36598
(±0.00337)
Rank: 5/7
0.30354
Rank: 6/7
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 20-01-28
ORB (OpenCV), DEGENSAC
Size: 32 bytes. Matches: built-in
7131.47 154.44
Rank: 6/7
0.506
Rank: 2/7
0.638
Rank: 7/7
0.15923
(±0.00080)
Rank: 7/7
902.07
Rank: 2/7
1303.19
Rank: 7/7
2.669
Rank: 7/7
0.911
Rank: 7/7
0.07230
(±0.00216)
Rank: 7/7
0.11577
Rank: 7/7
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 20-01-28
Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7874.85 260.65
Rank: 3/7
0.470
Rank: 5/7
0.830
Rank: 1/7
0.48001
(±0.00027)
Rank: 2/7
414.38
Rank: 6/7
3552.72
Rank: 3/7
4.032
Rank: 3/7
0.515
Rank: 2/7
0.59585
(±0.00212)
Rank: 3/7
0.53793
Rank: 3/7
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 20-01-28

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)
Authors Keypoint Descriptor Summary Paper Website Processing date
LogPolarDesc, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 144.16
Rank: 5/30
0.317
Rank: 17/30
0.740
Rank: 20/30
0.38507
(±0.00039)
Rank: 9/30
155.56
Rank: 18/30
1300.54
Rank: 9/30
3.954
Rank: 11/30
0.570
Rank: 8/30
0.52445
(±0.00315)
Rank: 8/30
0.45476
Rank: 9/30
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 20-01-28
GeoDesc (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 139.91
Rank: 6/30
0.327
Rank: 6/30
0.751
Rank: 14/30
0.37669
(±0.00080)
Rank: 14/30
141.85
Rank: 22/30
1155.98
Rank: 19/30
4.111
Rank: 8/30
0.591
Rank: 15/30
0.51746
(±0.00146)
Rank: 10/30
0.44708
Rank: 13/30
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 20-01-28
HardNet, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 105.71
Rank: 17/30
0.317
Rank: 17/30
0.765
Rank: 10/30
0.38298
(±0.00073)
Rank: 11/30
146.17
Rank: 20/30
1223.02
Rank: 14/30
3.934
Rank: 15/30
0.587
Rank: 14/30
0.50998
(±0.00237)
Rank: 15/30
0.44648
Rank: 14/30
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 20-01-28
R2D2 (r2d2-wasf-n16), DEGENSAC
Size: 512 bytes. Matches: built-in
2048.00 170.97
Rank: 2/30
0.418
Rank: 1/30
0.716
Rank: 23/30
0.35290
(±0.00069)
Rank: 19/30
253.03
Rank: 6/30
1134.39
Rank: 21/30
4.201
Rank: 3/30
0.546
Rank: 5/30
0.55946
(±0.00119)
Rank: 5/30
0.45618
Rank: 8/30
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 20-01-28
DELF-GLD (128D), PyRANSAC
Size: 512 bytes. Matches: built-in
2037.95 94.38
Rank: 22/30
0.110
Rank: 27/30
0.152
Rank: 27/30
0.06771
(±0.00064)
Rank: 28/30
422.01
Rank: 3/30
2016.18
Rank: 1/30
2.524
Rank: 26/30
0.908
Rank: 28/30
0.12029
(±0.00117)
Rank: 26/30
0.09400
Rank: 27/30
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 20-01-28
L2-Net, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 104.04
Rank: 18/30
0.317
Rank: 23/30
0.754
Rank: 13/30
0.35397
(±0.00088)
Rank: 18/30
132.80
Rank: 26/30
1129.03
Rank: 23/30
3.854
Rank: 20/30
0.592
Rank: 17/30
0.47583
(±0.00180)
Rank: 18/30
0.41490
Rank: 19/30
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 20-01-28
SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1939.53 74.14
Rank: 27/30
0.317
Rank: 15/30
0.721
Rank: 22/30
0.30329
(±0.00076)
Rank: 23/30
167.10
Rank: 14/30
986.36
Rank: 28/30
3.596
Rank: 23/30
0.671
Rank: 23/30
0.38429
(±0.00130)
Rank: 23/30
0.34379
Rank: 23/30
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 20-01-28
HardNet, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 134.14
Rank: 8/30
0.317
Rank: 17/30
0.746
Rank: 18/30
0.38203
(±0.00091)
Rank: 13/30
146.17
Rank: 20/30
1223.02
Rank: 14/30
3.934
Rank: 15/30
0.585
Rank: 13/30
0.50998
(±0.00237)
Rank: 15/30
0.44601
Rank: 15/30
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 20-01-28
L2-Net (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 133.44
Rank: 9/30
0.327
Rank: 6/30
0.772
Rank: 4/30
0.39481
(±0.00042)
Rank: 6/30
163.19
Rank: 15/30
1245.79
Rank: 12/30
4.115
Rank: 6/30
0.554
Rank: 6/30
0.53872
(±0.00257)
Rank: 6/30
0.46676
Rank: 6/30
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 20-01-28
DELF-GLD (128D), DEGENSAC
Size: 512 bytes. Matches: built-in
2037.95 100.52
Rank: 19/30
0.110
Rank: 27/30
0.148
Rank: 28/30
0.07695
(±0.00044)
Rank: 27/30
422.01
Rank: 3/30
2016.18
Rank: 1/30
2.524
Rank: 26/30
0.902
Rank: 27/30
0.12029
(±0.00117)
Rank: 26/30
0.09862
Rank: 26/30
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 20-01-28
ORB (OpenCV), DEGENSAC
Size: 32 bytes. Matches: built-in
2033.08 60.68
Rank: 30/30
0.356
Rank: 2/30
0.509
Rank: 26/30
0.08003
(±0.00082)
Rank: 26/30
119.41
Rank: 28/30
252.10
Rank: 30/30
1.974
Rank: 30/30
0.870
Rank: 26/30
0.02068
(±0.00199)
Rank: 30/30
0.05035
Rank: 30/30
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 20-01-28
SURF (OpenCV), DEGENSAC
Size: 256 bytes. Matches: built-in
2013.95 69.86
Rank: 28/30
0.319
Rank: 14/30
0.626
Rank: 24/30
0.20564
(±0.00049)
Rank: 24/30
294.61
Rank: 5/30
1130.23
Rank: 22/30
3.466
Rank: 24/30
0.738
Rank: 24/30
0.34357
(±0.00354)
Rank: 24/30
0.27460
Rank: 24/30
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 20-01-28
HardNet (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 175.29
Rank: 1/30
0.327
Rank: 6/30
0.768
Rank: 7/30
0.43287
(±0.00055)
Rank: 3/30
182.78
Rank: 10/30
1345.73
Rank: 5/30
4.201
Rank: 4/30
0.527
Rank: 1/30
0.57391
(±0.00317)
Rank: 1/30
0.50339
Rank: 2/30
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 20-01-28
GeoDesc, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 90.21
Rank: 24/30
0.317
Rank: 23/30
0.747
Rank: 17/30
0.34028
(±0.00040)
Rank: 20/30
114.97
Rank: 29/30
1018.30
Rank: 26/30
3.889
Rank: 18/30
0.608
Rank: 19/30
0.45972
(±0.00065)
Rank: 20/30
0.40000
Rank: 20/30
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 20-01-28
DELF-GLD (32D), PyRANSAC
Size: 128 bytes. Matches: built-in
2037.95 91.34
Rank: 23/30
0.110
Rank: 27/30
0.120
Rank: 29/30
0.05664
(±0.00023)
Rank: 30/30
444.19
Rank: 1/30
1680.65
Rank: 3/30
2.420
Rank: 28/30
0.942
Rank: 29/30
0.08292
(±0.00119)
Rank: 28/30
0.06978
Rank: 29/30
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 20-01-28
SOSNet (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 149.80
Rank: 4/30
0.327
Rank: 6/30
0.781
Rank: 3/30
0.42502
(±0.00011)
Rank: 4/30
173.88
Rank: 12/30
1312.03
Rank: 7/30
4.208
Rank: 1/30
0.529
Rank: 2/30
0.57242
(±0.00097)
Rank: 3/30
0.49872
Rank: 4/30
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 20-01-28
L2-Net (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 106.35
Rank: 16/30
0.327
Rank: 6/30
0.787
Rank: 2/30
0.40199
(±0.00065)
Rank: 5/30
163.19
Rank: 15/30
1245.79
Rank: 12/30
4.115
Rank: 6/30
0.556
Rank: 7/30
0.53872
(±0.00257)
Rank: 6/30
0.47035
Rank: 5/30
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 20-01-28
LogPolarDesc, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 112.92
Rank: 12/30
0.317
Rank: 17/30
0.759
Rank: 11/30
0.38939
(±0.00082)
Rank: 7/30
155.56
Rank: 18/30
1300.54
Rank: 9/30
3.954
Rank: 11/30
0.571
Rank: 9/30
0.52445
(±0.00315)
Rank: 8/30
0.45692
Rank: 7/30
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 20-01-28
SOSNet, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 115.88
Rank: 11/30
0.317
Rank: 17/30
0.757
Rank: 12/30
0.37316
(±0.00044)
Rank: 15/30
138.50
Rank: 24/30
1186.36
Rank: 17/30
3.953
Rank: 13/30
0.572
Rank: 10/30
0.51487
(±0.00216)
Rank: 12/30
0.44402
Rank: 16/30
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 20-01-28
Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1939.53 80.40
Rank: 26/30
0.317
Rank: 15/30
0.743
Rank: 19/30
0.33160
(±0.00070)
Rank: 22/30
162.54
Rank: 17/30
1064.73
Rank: 25/30
3.693
Rank: 22/30
0.633
Rank: 22/30
0.42461
(±0.00081)
Rank: 22/30
0.37811
Rank: 22/30
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 20-01-28
GeoDesc, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 111.61
Rank: 14/30
0.317
Rank: 23/30
0.731
Rank: 21/30
0.33818
(±0.00032)
Rank: 21/30
114.97
Rank: 29/30
1018.30
Rank: 26/30
3.889
Rank: 18/30
0.609
Rank: 20/30
0.45972
(±0.00065)
Rank: 20/30
0.39895
Rank: 21/30
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 20-01-28
SOSNet, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 130.42
Rank: 10/30
0.317
Rank: 17/30
0.748
Rank: 16/30
0.38360
(±0.00045)
Rank: 10/30
138.50
Rank: 24/30
1186.36
Rank: 17/30
3.953
Rank: 13/30
0.576
Rank: 12/30
0.51487
(±0.00216)
Rank: 12/30
0.44923
Rank: 12/30
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 20-01-28
Upright Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1893.01 107.42
Rank: 15/30
0.327
Rank: 4/30
0.768
Rank: 8/30
0.38929
(±0.00031)
Rank: 8/30
193.67
Rank: 9/30
1263.61
Rank: 11/30
3.995
Rank: 10/30
0.574
Rank: 11/30
0.51247
(±0.00292)
Rank: 14/30
0.45088
Rank: 10/30
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 20-01-28
DELF-GLD (32D), DEGENSAC
Size: 128 bytes. Matches: built-in
2037.95 96.84
Rank: 21/30
0.110
Rank: 27/30
0.117
Rank: 30/30
0.05987
(±0.00031)
Rank: 29/30
444.19
Rank: 1/30
1680.65
Rank: 3/30
2.420
Rank: 28/30
0.947
Rank: 30/30
0.08292
(±0.00119)
Rank: 28/30
0.07139
Rank: 28/30
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 20-01-28
Upright SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1893.01 100.40
Rank: 20/30
0.327
Rank: 4/30
0.750
Rank: 15/30
0.36487
(±0.00048)
Rank: 16/30
198.66
Rank: 8/30
1221.16
Rank: 16/30
3.908
Rank: 17/30
0.623
Rank: 21/30
0.47743
(±0.00204)
Rank: 17/30
0.42115
Rank: 17/30
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 20-01-28
HardNet (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 136.15
Rank: 7/30
0.327
Rank: 6/30
0.789
Rank: 1/30
0.43694
(±0.00116)
Rank: 1/30
182.78
Rank: 10/30
1345.73
Rank: 5/30
4.201
Rank: 4/30
0.529
Rank: 3/30
0.57391
(±0.00317)
Rank: 1/30
0.50543
Rank: 1/30
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 20-01-28
SOSNet (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 170.17
Rank: 3/30
0.327
Rank: 6/30
0.769
Rank: 6/30
0.43289
(±0.00101)
Rank: 2/30
173.88
Rank: 12/30
1312.03
Rank: 7/30
4.208
Rank: 1/30
0.531
Rank: 4/30
0.57242
(±0.00097)
Rank: 3/30
0.50266
Rank: 3/30
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 20-01-28
L2-Net, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 84.00
Rank: 25/30
0.317
Rank: 23/30
0.767
Rank: 9/30
0.36004
(±0.00079)
Rank: 17/30
132.80
Rank: 26/30
1129.03
Rank: 23/30
3.854
Rank: 20/30
0.591
Rank: 16/30
0.47583
(±0.00180)
Rank: 18/30
0.41793
Rank: 18/30
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 20-01-28
GeoDesc (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 112.15
Rank: 13/30
0.327
Rank: 6/30
0.770
Rank: 5/30
0.38207
(±0.00065)
Rank: 12/30
141.85
Rank: 22/30
1155.98
Rank: 19/30
4.111
Rank: 8/30
0.592
Rank: 18/30
0.51746
(±0.00146)
Rank: 10/30
0.44977
Rank: 11/30
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 20-01-28
AKAZE (OpenCV), DEGENSAC
Size: 61 bytes. Matches: built-in
1640.91 67.54
Rank: 29/30
0.352
Rank: 3/30
0.621
Rank: 25/30
0.17767
(±0.00068)
Rank: 25/30
219.35
Rank: 7/30
745.29
Rank: 29/30
3.279
Rank: 25/30
0.850
Rank: 25/30
0.21136
(±0.00413)
Rank: 25/30
0.19452
Rank: 25/30
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 20-01-28

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)
Authors Keypoint Descriptor Summary Paper Website Processing date
AKAZE (OpenCV), DEGENSAC
Size: 61 bytes. Matches: built-in
7871.55 236.04
Rank: 1/2
0.552
Rank: 1/2
0.723
Rank: 1/2
0.30219
(±0.00109)
Rank: 1/2
471.17
Rank: 2/2
2585.81
Rank: 1/2
3.327
Rank: 1/2
0.772
Rank: 1/2
0.32662
(±0.00278)
Rank: 1/2
0.31441
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 20-01-28
ORB (OpenCV), DEGENSAC
Size: 32 bytes. Matches: built-in
7131.47 154.44
Rank: 2/2
0.506
Rank: 2/2
0.638
Rank: 2/2
0.15923
(±0.00080)
Rank: 2/2
902.07
Rank: 1/2
1303.19
Rank: 2/2
2.669
Rank: 2/2
0.911
Rank: 2/2
0.07230
(±0.00216)
Rank: 2/2
0.11577
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 20-01-28

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)
Authors Keypoint Descriptor Summary Paper Website Processing date
ORB (OpenCV), DEGENSAC
Size: 32 bytes. Matches: built-in
2033.08 60.68
Rank: 4/4
0.356
Rank: 1/4
0.509
Rank: 2/4
0.08003
(±0.00082)
Rank: 2/4
119.41
Rank: 4/4
252.10
Rank: 4/4
1.974
Rank: 4/4
0.870
Rank: 2/4
0.02068
(±0.00199)
Rank: 4/4
0.05035
Rank: 4/4
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 20-01-28
DELF-GLD (32D), PyRANSAC
Size: 128 bytes. Matches: built-in
2037.95 91.34
Rank: 2/4
0.110
Rank: 3/4
0.120
Rank: 3/4
0.05664
(±0.00023)
Rank: 4/4
444.19
Rank: 1/4
1680.65
Rank: 1/4
2.420
Rank: 2/4
0.942
Rank: 3/4
0.08292
(±0.00119)
Rank: 2/4
0.06978
Rank: 3/4
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 20-01-28
DELF-GLD (32D), DEGENSAC
Size: 128 bytes. Matches: built-in
2037.95 96.84
Rank: 1/4
0.110
Rank: 3/4
0.117
Rank: 4/4
0.05987
(±0.00031)
Rank: 3/4
444.19
Rank: 1/4
1680.65
Rank: 1/4
2.420
Rank: 2/4
0.947
Rank: 4/4
0.08292
(±0.00119)
Rank: 2/4
0.07139
Rank: 2/4
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 20-01-28
AKAZE (OpenCV), DEGENSAC
Size: 61 bytes. Matches: built-in
1640.91 67.54
Rank: 3/4
0.352
Rank: 2/4
0.621
Rank: 1/4
0.17767
(±0.00068)
Rank: 1/4
219.35
Rank: 3/4
745.29
Rank: 3/4
3.279
Rank: 1/4
0.850
Rank: 1/4
0.21136
(±0.00413)
Rank: 1/4
0.19452
Rank: 1/4
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 20-01-28

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)
Authors Keypoint Descriptor Summary Paper Website Processing date
SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7874.85 226.69
Rank: 7/9
0.470
Rank: 5/9
0.809
Rank: 4/9
0.44473
(±0.00120)
Rank: 4/9
399.54
Rank: 9/9
3265.87
Rank: 6/9
3.883
Rank: 4/9
0.561
Rank: 4/9
0.54747
(±0.00225)
Rank: 4/9
0.49610
Rank: 4/9
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 20-01-28
Upright Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7846.57 336.38
Rank: 1/9
0.485
Rank: 3/9
0.830
Rank: 2/9
0.50131
(±0.00017)
Rank: 1/9
516.33
Rank: 5/9
4113.38
Rank: 3/9
4.260
Rank: 1/9
0.490
Rank: 1/9
0.63342
(±0.00110)
Rank: 1/9
0.56737
Rank: 1/9
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 20-01-28
Upright SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7846.57 302.54
Rank: 3/9
0.485
Rank: 3/9
0.815
Rank: 3/9
0.47840
(±0.00047)
Rank: 3/9
499.07
Rank: 6/9
3865.68
Rank: 4/9
4.139
Rank: 2/9
0.524
Rank: 3/9
0.60037
(±0.00279)
Rank: 2/9
0.53938
Rank: 2/9
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 20-01-28
AKAZE (OpenCV), DEGENSAC
Size: 61 bytes. Matches: built-in
7871.55 236.04
Rank: 6/9
0.552
Rank: 1/9
0.723
Rank: 5/9
0.30219
(±0.00109)
Rank: 5/9
471.17
Rank: 7/9
2585.81
Rank: 8/9
3.327
Rank: 7/9
0.772
Rank: 7/9
0.32662
(±0.00278)
Rank: 8/9
0.31441
Rank: 5/9
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 20-01-28
D2-Net (multi-scale), DEGENSAC
Size: 2048 bytes. Matches: built-in
6882.73 329.08
Rank: 2/9
0.319
Rank: 8/9
0.374
Rank: 8/9
0.19670
(±0.00066)
Rank: 7/9
712.75
Rank: 4/9
5309.21
Rank: 1/9
3.302
Rank: 8/9
0.689
Rank: 5/9
0.38916
(±0.00459)
Rank: 5/9
0.29293
Rank: 7/9
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 20-01-28
SURF (OpenCV), DEGENSAC
Size: 256 bytes. Matches: built-in
7750.71 121.14
Rank: 9/9
0.433
Rank: 7/9
0.667
Rank: 6/9
0.24110
(±0.00061)
Rank: 6/9
913.30
Rank: 1/9
3217.44
Rank: 7/9
3.353
Rank: 6/9
0.780
Rank: 8/9
0.36598
(±0.00337)
Rank: 7/9
0.30354
Rank: 6/9
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 20-01-28
D2-Net (single-scale), DEGENSAC
Size: 2048 bytes. Matches: built-in
5598.19 242.31
Rank: 5/9
0.285
Rank: 9/9
0.352
Rank: 9/9
0.17876
(±0.00032)
Rank: 8/9
714.41
Rank: 3/9
4274.86
Rank: 2/9
3.456
Rank: 5/9
0.730
Rank: 6/9
0.38157
(±0.00162)
Rank: 6/9
0.28017
Rank: 8/9
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 20-01-28
ORB (OpenCV), DEGENSAC
Size: 32 bytes. Matches: built-in
7131.47 154.44
Rank: 8/9
0.506
Rank: 2/9
0.638
Rank: 7/9
0.15923
(±0.00080)
Rank: 9/9
902.07
Rank: 2/9
1303.19
Rank: 9/9
2.669
Rank: 9/9
0.911
Rank: 9/9
0.07230
(±0.00216)
Rank: 9/9
0.11577
Rank: 9/9
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 20-01-28
Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
7874.85 260.65
Rank: 4/9
0.470
Rank: 5/9
0.830
Rank: 1/9
0.48001
(±0.00027)
Rank: 2/9
414.38
Rank: 8/9
3552.72
Rank: 5/9
4.032
Rank: 3/9
0.515
Rank: 2/9
0.59585
(±0.00212)
Rank: 3/9
0.53793
Rank: 3/9
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 20-01-28

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)
Authors Keypoint Descriptor Summary Paper Website Processing date
SuperPoint (2k features, NMS=4), DEGENSA...
Size: 1024 bytes. Matches: built-in
2048.00 117.93
Rank: 12/37
0.359
Rank: 4/37
0.621
Rank: 26/37
0.28007
(±0.00041)
Rank: 24/37
161.96
Rank: 22/37
1097.54
Rank: 29/37
4.211
Rank: 1/37
0.611
Rank: 21/37
0.49880
(±0.00310)
Rank: 17/37
0.38943
Rank: 22/37
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 20-01-28
LogPolarDesc, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 144.16
Rank: 5/37
0.317
Rank: 20/37
0.740
Rank: 20/37
0.38507
(±0.00039)
Rank: 9/37
155.56
Rank: 23/37
1300.54
Rank: 13/37
3.954
Rank: 14/37
0.570
Rank: 8/37
0.52445
(±0.00315)
Rank: 8/37
0.45476
Rank: 9/37
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 20-01-28
GeoDesc (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 139.91
Rank: 7/37
0.327
Rank: 9/37
0.751
Rank: 14/37
0.37669
(±0.00080)
Rank: 14/37
141.85
Rank: 28/37
1155.98
Rank: 23/37
4.111
Rank: 10/37
0.591
Rank: 15/37
0.51746
(±0.00146)
Rank: 10/37
0.44708
Rank: 13/37
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 20-01-28
D2-Net (single-scale), DEGENSAC
Size: 2048 bytes. Matches: built-in
2039.21 144.14
Rank: 6/37
0.199
Rank: 30/37
0.332
Rank: 30/37
0.15679
(±0.00080)
Rank: 29/37
355.63
Rank: 7/37
1859.99
Rank: 6/37
3.382
Rank: 28/37
0.756
Rank: 28/37
0.33547
(±0.00111)
Rank: 28/37
0.24613
Rank: 28/37
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 20-01-28
HardNet, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 105.71
Rank: 21/37
0.317
Rank: 20/37
0.765
Rank: 10/37
0.38298
(±0.00073)
Rank: 11/37
146.17
Rank: 26/37
1223.02
Rank: 18/37
3.934
Rank: 18/37
0.587
Rank: 14/37
0.50998
(±0.00237)
Rank: 15/37
0.44648
Rank: 14/37
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 20-01-28
R2D2 (r2d2-wasf-n16), DEGENSAC
Size: 512 bytes. Matches: built-in
2048.00 170.97
Rank: 2/37
0.418
Rank: 1/37
0.716
Rank: 23/37
0.35290
(±0.00069)
Rank: 19/37
253.03
Rank: 10/37
1134.39
Rank: 25/37
4.201
Rank: 4/37
0.546
Rank: 5/37
0.55946
(±0.00119)
Rank: 5/37
0.45618
Rank: 8/37
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 20-01-28
DELF-GLD (128D), PyRANSAC
Size: 512 bytes. Matches: built-in
2037.95 94.38
Rank: 28/37
0.110
Rank: 32/37
0.152
Rank: 34/37
0.06771
(±0.00064)
Rank: 35/37
422.01
Rank: 5/37
2016.18
Rank: 3/37
2.524
Rank: 33/37
0.908
Rank: 35/37
0.12029
(±0.00117)
Rank: 33/37
0.09400
Rank: 34/37
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 20-01-28
L2-Net, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 104.04
Rank: 22/37
0.317
Rank: 26/37
0.754
Rank: 13/37
0.35397
(±0.00088)
Rank: 18/37
132.80
Rank: 32/37
1129.03
Rank: 27/37
3.854
Rank: 23/37
0.592
Rank: 17/37
0.47583
(±0.00180)
Rank: 20/37
0.41490
Rank: 19/37
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 20-01-28
SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1939.53 74.14
Rank: 34/37
0.317
Rank: 18/37
0.721
Rank: 22/37
0.30329
(±0.00076)
Rank: 23/37
167.10
Rank: 18/37
986.36
Rank: 34/37
3.596
Rank: 26/37
0.671
Rank: 26/37
0.38429
(±0.00130)
Rank: 26/37
0.34379
Rank: 26/37
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 20-01-28
SuperPoint (2k features, NMS=2), DEGENSA...
Size: 1024 bytes. Matches: built-in
2048.00 101.75
Rank: 23/37
0.371
Rank: 2/37
0.615
Rank: 28/37
0.27040
(±0.00072)
Rank: 26/37
124.15
Rank: 34/37
928.73
Rank: 35/37
4.094
Rank: 12/37
0.643
Rank: 25/37
0.46511
(±0.00392)
Rank: 22/37
0.36776
Rank: 25/37
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 20-01-28
HardNet, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 134.14
Rank: 9/37
0.317
Rank: 20/37
0.746
Rank: 18/37
0.38203
(±0.00091)
Rank: 13/37
146.17
Rank: 26/37
1223.02
Rank: 18/37
3.934
Rank: 18/37
0.585
Rank: 13/37
0.50998
(±0.00237)
Rank: 15/37
0.44601
Rank: 15/37
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 20-01-28
D2-Net (multi-scale), DEGENSAC
Size: 2048 bytes. Matches: built-in
2045.80 117.10
Rank: 13/37
0.167
Rank: 31/37
0.287
Rank: 31/37
0.11872
(±0.00058)
Rank: 30/37
277.73
Rank: 9/37
1879.74
Rank: 5/37
3.020
Rank: 30/37
0.798
Rank: 29/37
0.25757
(±0.00178)
Rank: 29/37
0.18815
Rank: 30/37
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 20-01-28
SuperPoint (2k features, NMS=3), DEGENSA...
Size: 1024 bytes. Matches: built-in
2048.00 112.71
Rank: 16/37
0.364
Rank: 3/37
0.619
Rank: 27/37
0.27741
(±0.00074)
Rank: 25/37
148.24
Rank: 25/37
1037.77
Rank: 31/37
4.195
Rank: 7/37
0.621
Rank: 22/37
0.49001
(±0.00124)
Rank: 18/37
0.38371
Rank: 23/37
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 20-01-28
L2-Net (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 133.44
Rank: 10/37
0.327
Rank: 9/37
0.772
Rank: 4/37
0.39481
(±0.00042)
Rank: 6/37
163.19
Rank: 19/37
1245.79
Rank: 16/37
4.115
Rank: 8/37
0.554
Rank: 6/37
0.53872
(±0.00257)
Rank: 6/37
0.46676
Rank: 6/37
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 20-01-28
DELF-GLD (128D), DEGENSAC
Size: 512 bytes. Matches: built-in
2037.95 100.52
Rank: 24/37
0.110
Rank: 32/37
0.148
Rank: 35/37
0.07695
(±0.00044)
Rank: 33/37
422.01
Rank: 5/37
2016.18
Rank: 3/37
2.524
Rank: 33/37
0.902
Rank: 33/37
0.12029
(±0.00117)
Rank: 33/37
0.09862
Rank: 33/37
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 20-01-28
ORB (OpenCV), DEGENSAC
Size: 32 bytes. Matches: built-in
2033.08 60.68
Rank: 37/37
0.356
Rank: 5/37
0.509
Rank: 29/37
0.08003
(±0.00082)
Rank: 32/37
119.41
Rank: 35/37
252.10
Rank: 37/37
1.974
Rank: 37/37
0.870
Rank: 31/37
0.02068
(±0.00199)
Rank: 37/37
0.05035
Rank: 37/37
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 20-01-28
SURF (OpenCV), DEGENSAC
Size: 256 bytes. Matches: built-in
2013.95 69.86
Rank: 35/37
0.319
Rank: 17/37
0.626
Rank: 24/37
0.20564
(±0.00049)
Rank: 27/37
294.61
Rank: 8/37
1130.23
Rank: 26/37
3.466
Rank: 27/37
0.738
Rank: 27/37
0.34357
(±0.00354)
Rank: 27/37
0.27460
Rank: 27/37
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 20-01-28
HardNet (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 175.29
Rank: 1/37
0.327
Rank: 9/37
0.768
Rank: 7/37
0.43287
(±0.00055)
Rank: 3/37
182.78
Rank: 14/37
1345.73
Rank: 9/37
4.201
Rank: 5/37
0.527
Rank: 1/37
0.57391
(±0.00317)
Rank: 1/37
0.50339
Rank: 2/37
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 20-01-28
GeoDesc, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 90.21
Rank: 31/37
0.317
Rank: 26/37
0.747
Rank: 17/37
0.34028
(±0.00040)
Rank: 20/37
114.97
Rank: 36/37
1018.30
Rank: 32/37
3.889
Rank: 21/37
0.608
Rank: 19/37
0.45972
(±0.00065)
Rank: 23/37
0.40000
Rank: 20/37
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 20-01-28
DELF-GLD (32D), PyRANSAC
Size: 128 bytes. Matches: built-in
2037.95 91.34
Rank: 30/37
0.110
Rank: 32/37
0.120
Rank: 36/37
0.05664
(±0.00023)
Rank: 37/37
444.19
Rank: 1/37
1680.65
Rank: 7/37
2.420
Rank: 35/37
0.942
Rank: 36/37
0.08292
(±0.00119)
Rank: 35/37
0.06978
Rank: 36/37
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 20-01-28
SOSNet (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 149.80
Rank: 4/37
0.327
Rank: 9/37
0.781
Rank: 3/37
0.42502
(±0.00011)
Rank: 4/37
173.88
Rank: 16/37
1312.03
Rank: 11/37
4.208
Rank: 2/37
0.529
Rank: 2/37
0.57242
(±0.00097)
Rank: 3/37
0.49872
Rank: 4/37
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 20-01-28
L2-Net (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 106.35
Rank: 20/37
0.327
Rank: 9/37
0.787
Rank: 2/37
0.40199
(±0.00065)
Rank: 5/37
163.19
Rank: 19/37
1245.79
Rank: 16/37
4.115
Rank: 8/37
0.556
Rank: 7/37
0.53872
(±0.00257)
Rank: 6/37
0.47035
Rank: 5/37
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 20-01-28
LogPolarDesc, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 112.92
Rank: 15/37
0.317
Rank: 20/37
0.759
Rank: 11/37
0.38939
(±0.00082)
Rank: 7/37
155.56
Rank: 23/37
1300.54
Rank: 13/37
3.954
Rank: 14/37
0.571
Rank: 9/37
0.52445
(±0.00315)
Rank: 8/37
0.45692
Rank: 7/37
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 20-01-28
SOSNet, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 115.88
Rank: 14/37
0.317
Rank: 20/37
0.757
Rank: 12/37
0.37316
(±0.00044)
Rank: 15/37
138.50
Rank: 30/37
1186.36
Rank: 21/37
3.953
Rank: 16/37
0.572
Rank: 10/37
0.51487
(±0.00216)
Rank: 12/37
0.44402
Rank: 16/37
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 20-01-28
Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1939.53 80.40
Rank: 33/37
0.317
Rank: 18/37
0.743
Rank: 19/37
0.33160
(±0.00070)
Rank: 22/37
162.54
Rank: 21/37
1064.73
Rank: 30/37
3.693
Rank: 25/37
0.633
Rank: 24/37
0.42461
(±0.00081)
Rank: 25/37
0.37811
Rank: 24/37
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 20-01-28
DELF-GLD (512D), PyRANSAC
Size: 2048 bytes. Matches: built-in
2037.95 92.64
Rank: 29/37
0.110
Rank: 32/37
0.171
Rank: 32/37
0.07179
(±0.00016)
Rank: 34/37
429.05
Rank: 3/37
2136.19
Rank: 1/37
2.554
Rank: 31/37
0.906
Rank: 34/37
0.13580
(±0.00450)
Rank: 31/37
0.10379
Rank: 32/37
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 20-01-28
GeoDesc, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 111.61
Rank: 18/37
0.317
Rank: 26/37
0.731
Rank: 21/37
0.33818
(±0.00032)
Rank: 21/37
114.97
Rank: 36/37
1018.30
Rank: 32/37
3.889
Rank: 21/37
0.609
Rank: 20/37
0.45972
(±0.00065)
Rank: 23/37
0.39895
Rank: 21/37
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 20-01-28
SOSNet, MAGSAC
Size: 512 bytes. Matches: built-in
1939.51 130.42
Rank: 11/37
0.317
Rank: 20/37
0.748
Rank: 16/37
0.38360
(±0.00045)
Rank: 10/37
138.50
Rank: 30/37
1186.36
Rank: 21/37
3.953
Rank: 16/37
0.576
Rank: 12/37
0.51487
(±0.00216)
Rank: 12/37
0.44923
Rank: 12/37
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 20-01-28
Upright Root-SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1893.01 107.42
Rank: 19/37
0.327
Rank: 7/37
0.768
Rank: 8/37
0.38929
(±0.00031)
Rank: 8/37
193.67
Rank: 13/37
1263.61
Rank: 15/37
3.995
Rank: 13/37
0.574
Rank: 11/37
0.51247
(±0.00292)
Rank: 14/37
0.45088
Rank: 10/37
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 20-01-28
DELF-GLD (32D), DEGENSAC
Size: 128 bytes. Matches: built-in
2037.95 96.84
Rank: 27/37
0.110
Rank: 32/37
0.117
Rank: 37/37
0.05987
(±0.00031)
Rank: 36/37
444.19
Rank: 1/37
1680.65
Rank: 7/37
2.420
Rank: 35/37
0.947
Rank: 37/37
0.08292
(±0.00119)
Rank: 35/37
0.07139
Rank: 35/37
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 20-01-28
Upright SIFT (OpenCV), DEGENSAC
Size: 512 bytes. Matches: built-in
1893.01 100.40
Rank: 25/37
0.327
Rank: 7/37
0.750
Rank: 15/37
0.36487
(±0.00048)
Rank: 16/37
198.66
Rank: 12/37
1221.16
Rank: 20/37
3.908
Rank: 20/37
0.623
Rank: 23/37
0.47743
(±0.00204)
Rank: 19/37
0.42115
Rank: 17/37
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 20-01-28
HardNet (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 136.15
Rank: 8/37
0.327
Rank: 9/37
0.789
Rank: 1/37
0.43694
(±0.00116)
Rank: 1/37
182.78
Rank: 14/37
1345.73
Rank: 9/37
4.201
Rank: 5/37
0.529
Rank: 3/37
0.57391
(±0.00317)
Rank: 1/37
0.50543
Rank: 1/37
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 20-01-28
SOSNet (upright), MAGSAC
Size: 512 bytes. Matches: built-in
1892.99 170.17
Rank: 3/37
0.327
Rank: 9/37
0.769
Rank: 6/37
0.43289
(±0.00101)
Rank: 2/37
173.88
Rank: 16/37
1312.03
Rank: 11/37
4.208
Rank: 2/37
0.531
Rank: 4/37
0.57242
(±0.00097)
Rank: 3/37
0.50266
Rank: 3/37
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 20-01-28
L2-Net, DEGENSAC
Size: 512 bytes. Matches: built-in
1939.51 84.00
Rank: 32/37
0.317
Rank: 26/37
0.767
Rank: 9/37
0.36004
(±0.00079)
Rank: 17/37
132.80
Rank: 32/37
1129.03
Rank: 27/37
3.854
Rank: 23/37
0.591
Rank: 16/37
0.47583
(±0.00180)
Rank: 20/37
0.41793
Rank: 18/37
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 20-01-28
GeoDesc (upright), DEGENSAC
Size: 512 bytes. Matches: built-in
1892.99 112.15
Rank: 17/37
0.327
Rank: 9/37
0.770
Rank: 5/37
0.38207
(±0.00065)
Rank: 12/37
141.85
Rank: 28/37
1155.98
Rank: 23/37
4.111
Rank: 10/37
0.592
Rank: 18/37
0.51746
(±0.00146)
Rank: 10/37
0.44977
Rank: 11/37
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 20-01-28
AKAZE (OpenCV), DEGENSAC
Size: 61 bytes. Matches: built-in
1640.91 67.54
Rank: 36/37
0.352
Rank: 6/37
0.621
Rank: 25/37
0.17767
(±0.00068)
Rank: 28/37
219.35
Rank: 11/37
745.29
Rank: 36/37
3.279
Rank: 29/37
0.850
Rank: 30/37
0.21136
(±0.00413)
Rank: 30/37
0.19452
Rank: 29/37
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 20-01-28
DELF-GLD (512D), DEGENSAC
Size: 2048 bytes. Matches: built-in
2037.95 98.59
Rank: 26/37
0.110
Rank: 32/37
0.166
Rank: 33/37
0.08477
(±0.00027)
Rank: 31/37
429.05
Rank: 3/37
2136.19
Rank: 1/37
2.554
Rank: 31/37
0.899
Rank: 32/37
0.13580
(±0.00450)
Rank: 31/37
0.11028
Rank: 31/37
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 20-01-28