Breakdown: Phototourism | MVS | Subsets

Breakdown on the Phototourism dataset, multi-view stereo task, by subset size.

MVS, bag size 3
Method Date Type Ims (%) #Pts SR TL mAP5o mAP10o mAP15o mAP20o mAP25o ATE By Details Link Contact Updated Descriptor size
AKAZE (OpenCV)
kp:8000, match:nn		 19-04-24 F 100.0 692.9 57.5 2.28 0.0418 0.0885 0.1321 0.1794 0.2145 — Challenge organizers AKAZE, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SphereDesc
kp:8000, match:nn		 19-04-26 F 100.0 727.0 77.6 2.28 0.0785 0.1542 0.2267 0.2839 0.3364 — Anonymous We use OpenCV's implementation of AKAZE detector, and for each keypoint we extract a descriptor via CNN. N/A Anonymous N/A 256 float32
Brisk + SSS
kp:8000, match:nn		 19-05-14 F 100.0 424.4 58.2 2.16 0.0115 0.0409 0.0803 0.1176 0.1564 — Anonymous We use OpenCV's implementation of brisk detector with the default settings, and for each image there are at most 8K keypoints. For each keypoint, we extract a descriptor via a CNN model. TBA Anonymous N/A 128 float32
D2-Net (single scale)
kp:8000, match:nn		 19-05-07 F 100.0 752.7 72.0 2.23 0.0536 0.1255 0.1985 0.2521 0.2973 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multiscale)
kp:8000, match:nn		 19-05-07 F 100.0 896.8 78.2 2.20 0.0545 0.1324 0.1961 0.2667 0.3279 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (single scale, no PT dataset)
kp:8000, match:nn		 19-06-01 F 100.0 763.3 72.0 2.22 0.0488 0.1221 0.1870 0.2424 0.2961 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multi-scale, no PT dataset)
kp:8000, match:nn		 19-06-05 F 100.0 813.4 76.7 2.21 0.0418 0.1185 0.1939 0.2570 0.3194 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
DELF
kp:1024, match:nn		 19-05-05 F 100.0 143.7 32.4 2.12 0.0048 0.0233 0.0388 0.0603 0.0779 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:2048, match:nn		 19-05-05 F 100.0 226.5 47.5 2.10 0.0048 0.0324 0.0648 0.0997 0.1306 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:256, match:nn		 19-05-05 F 36.4 23.6 0.5 0.79 0.0000 0.0003 0.0003 0.0006 0.0012 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:512, match:nn		 19-05-05 F 90.9 75.0 10.3 2.00 0.0012 0.0058 0.0121 0.0164 0.0221 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
ELF-256D
kp:512, match:nn		 19-05-07 F 63.6 54.8 2.5 1.48 0.0021 0.0036 0.0061 0.0079 0.0103 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-512D
kp:512, match:nn		 19-05-09 F 72.7 57.9 4.9 1.65 0.0027 0.0058 0.0097 0.0124 0.0158 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-SIFT
kp:512, match:nn		 19-04-26 F 90.9 96.5 9.9 2.12 0.0073 0.0227 0.0303 0.0388 0.0476 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are HOG (as in SIFT). N/A Anonymous N/A 128 uint8
SIFT + GeoDesc
kp:2048, match:nn		 19-05-19 F 100.0 286.7 59.0 2.30 0.0439 0.1018 0.1470 0.1894 0.2252 — Challenge organizers GeoDesc extracted on SIFT keypoints. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:2048, match:nn		 19-05-19 F 100.0 291.2 57.2 2.29 0.0418 0.0927 0.1464 0.1885 0.2261 — Challenge organizers HardNet extracted on SIFT keypoints. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger patches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
HarrisZ/RsGLOH2
kp:8000, match:sGOr2f		 19-05-23 F 100.0 353.3 76.5 2.32 0.0821 0.1603 0.2273 0.2903 0.3376 — Fabio Bellavia and Carlo Colombo HarrisZ keypoints and Root squared (like RootSIFT) sGLOH2 descriptors using sGOr2f* matching strategy. Distance table entries lower than their respective row-wise and column-wise averages are discarded and then matching pairs are computed using the greedy nearest neighbour as in the WISW@CAIP2019 contest. Keypoints and matches are ordered according to their ranks (best ones first). http://cvg.dsi.unifi.it/cvg/index.php?id=research#descriptor bellavia.fabio@gmail.com N/A 256 float32
HesAffNet - HardNet2
kp:8000, match:nn		 19-05-29 F 100.0 1084.3 83.0 2.31 0.0973 0.1842 0.2561 0.3167 0.3776 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian + AffNet(affine shape) + OriNet (orientation), Code and weight from https://github.com/ducha-aiki/affnet, Paper: https://arxiv.org/abs/1711.06704; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset, Paper: https://arxiv.org/pdf/1901.09780.pdf; DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
Hessian - HardNet2
kp:8000, match:nn		 19-05-30 F 100.0 1259.3 77.3 2.34 0.1239 0.2027 0.2667 0.3270 0.3745 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian. Gravity vector orientation is assumed; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset Paper: https://arxiv.org/pdf/1901.09780.pdf, DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
ORB (OpenCV)
kp:8000, match:nn		 19-04-24 F 100.0 690.1 35.1 2.23 0.0112 0.0315 0.0552 0.0770 0.1000 — Challenge organizers ORB, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
Scale-invariant desc. (Log-Polar, lambda=32)
kp:8000, match:nn		 19-06-25 F 100.0 1084.1 86.0 2.27 0.0958 0.1836 0.2558 0.3167 0.3752 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=64)
kp:8000, match:nn		 19-06-24 F 100.0 1149.0 85.6 2.27 0.0979 0.1842 0.2588 0.3121 0.3636 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=96)
kp:8000, match:nn		 19-06-20 F 100.0 1139.8 88.0 2.26 0.0967 0.1824 0.2573 0.3200 0.3727 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SIFT-AID (NN matcher)
kp:8000, match:nn		 19-05-10 F 100.0 432.1 38.7 2.19 0.0100 0.0339 0.0597 0.0842 0.1130 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT-AID (custom matcher)
kp:8000, match:sift-aid		 19-04-29 F/M 100.0 389.7 35.5 2.19 0.0082 0.0300 0.0558 0.0794 0.1048 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT + ContextDesc
kp:8000, match:nn		 19-05-09 F 100.0 992.1 88.9 2.26 0.1033 0.1945 0.2542 0.3109 0.3618 — Zixin Luo ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 uint8
SIFT-Dense-ContextDesc
kp:8000, match:nn		 19-05-28 F 100.0 923.7 79.6 2.26 0.0924 0.1636 0.2297 0.2855 0.3291 — Zixin Luo, Jiahui Zhang Dense-ContextDesc is a variant of ContextDesc, where descriptors are densely extracted from full images, instead of image patches, while other settings stay unchanged as original ContextDesc. We find Dense-ContextDesc performs better regarding in particular illumination changes. Dense-ContextDesc is extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. The model will be available on the authors' GitHub page. TBA zluoag@cse.ust.hk N/A TBA
SIFT + ContextDesc + Inlier Classification V2
kp:8000, match:custom		 19-05-28 F/M 100.0 1021.4 90.6 2.32 0.2618 0.4164 0.5167 0.5803 0.6297 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an improved inlier classification and fundamental matrix estimation network based on [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT-GeoDesc-GitHub
kp:8000, match:nn		 19-05-08 F 100.0 804.8 82.5 2.28 0.0924 0.1691 0.2330 0.2900 0.3494 — Zixin Luo GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features extracted from the code provided by the authors. https://arxiv.org/abs/1807.06294 zluoag@cse.ust.hk N/A 128 float32
SIFT + GeoDesc
kp:8000, match:nn		 19-04-24 F 100.0 880.0 83.5 2.30 0.0906 0.1715 0.2261 0.2915 0.3382 — Challenge organizers GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:8000, match:nn		 19-04-24 F 100.0 1073.9 88.7 2.29 0.1124 0.2042 0.2739 0.3312 0.3842 — Challenge organizers HardNet extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
SIFT + L2-Net
kp:8000, match:nn		 19-04-24 F 100.0 948.2 85.5 2.26 0.0797 0.1606 0.2309 0.2836 0.3336 — Challenge organizers L2-Net extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/yuruntian/L2-Net imagematching@uvic.ca N/A 128 float32
SIFT + ContextDesc + Inlier Classification V1
kp:8000, match:custom		 19-05-29 F/M 100.0 992.7 91.6 2.30 0.2548 0.3994 0.4861 0.5518 0.6097 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an inlier classification and fundamental matrix estimation network using the architecture of [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT (OpenCV)
kp:8000, match:nn		 19-04-24 F 100.0 633.2 57.5 2.25 0.0470 0.0882 0.1273 0.1648 0.2018 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
SIFT + TFeat
kp:8000, match:nn		 19-04-24 F 100.0 801.8 77.3 2.24 0.0621 0.1233 0.1788 0.2342 0.2797 — Challenge organizers T-Feat extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/vbalnt/tfeat imagematching@uvic.ca N/A 128 float32
SIFT (OpenCV)
kp:2048, match:nn		 19-05-17 F 100.0 263.5 32.0 2.27 0.0145 0.0376 0.0588 0.0833 0.1024 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
Superpoint (nn matcher)
kp:2048, match:nn		 19-06-07 F 100.0 323.9 63.4 2.37 0.0700 0.1452 0.1967 0.2373 0.2806 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
Superpoint (1:1 matcher)
kp:2048, match:nn1to1		 19-06-07 F 100.0 308.6 67.5 2.40 0.1218 0.2124 0.2724 0.3230 0.3624 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
SuperPoint (default)
kp:2048, match:nn		 19-04-24 F 100.0 228.2 50.6 2.35 0.0527 0.1191 0.1694 0.2130 0.2464 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned (about 1200 on average). Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:1024, match:nn		 19-04-26 F 100.0 199.7 46.5 2.35 0.0530 0.1067 0.1488 0.1833 0.2145 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:2048, match:nn		 19-04-26 F 100.0 318.0 56.1 2.32 0.0488 0.1039 0.1576 0.1970 0.2321 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:256, match:nn		 19-04-26 F 54.5 43.5 6.5 1.34 0.0082 0.0164 0.0224 0.0239 0.0270 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:512, match:nn		 19-04-26 F 90.9 106.1 27.9 2.17 0.0412 0.0779 0.1003 0.1188 0.1333 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:8000, match:nn		 19-04-26 F 100.0 897.1 69.0 2.26 0.0403 0.1033 0.1630 0.2130 0.2612 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
Scale-invariant desc. (Cartesian, lambda=16)
kp:8000, match:nn		 19-07-29 F 100.0 1106.2 89.5 2.30 0.0964 0.1955 0.2724 0.3342 0.3836 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. (This is a baseline where we use cartesian patches instead.) Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SuperPoint (trained on coco + phototourism training set)
kp:2048, match:nn		 19-05-30 F 100.0 249.3 57.0 2.35 0.0533 0.1112 0.1600 0.2058 0.2442 — Daniel DeTone, Paul Sarlin, Tomasz Malisiewicz, Andrew Rabinovich SuperPoint V1 model trained on COCO homographic warps at VGA resolution, plus pairs from the phototourism training set using the GT poses and depths for correspondence. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork ddetone@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v1)
kp:2048, match:custom		 19-05-30 F/M 100.0 235.7 69.8 2.42 0.1621 0.2791 0.3542 0.4139 0.4582 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set; - keypoints refinement; - better descriptor sampling; - adjusted thresholds. A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v2)
kp:2048, match:custom		 19-05-28 F/M 100.0 284.3 69.8 2.39 0.1788 0.2912 0.3670 0.4252 0.4703 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set, - keypoints refinement, - better descriptor sampling, - adjusted thresholds; A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SURF (OpenCV)
kp:8000, match:nn		 19-04-24 F 100.0 455.0 37.4 2.20 0.0142 0.0370 0.0633 0.0855 0.1136 — Challenge organizers SURF, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SIFT + ContextDesc
kp:8000, match:nn1to1		 19-06-07 F 100.0 999.6 92.3 2.31 0.1597 0.2676 0.3542 0.4170 0.4694 — Challenge organizers ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with nearest-neighbour search, enforcing cross-match consistency. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc imagematching@uvic.ca N/A TBA
MVS, bag size 5
Method Date Type Ims (%) #Pts SR TL mAP5o mAP10o mAP15o mAP20o mAP25o ATE By Details Link Contact Updated Descriptor size
AKAZE (OpenCV)
kp:8000, match:nn		 19-04-24 F 88.0 1106.8 96.6 2.58 0.1263 0.2093 0.2677 0.3189 0.3642 — Challenge organizers AKAZE, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SphereDesc
kp:8000, match:nn		 19-04-26 F 92.7 1336.3 98.7 2.59 0.2007 0.3067 0.3751 0.4305 0.4774 — Anonymous We use OpenCV's implementation of AKAZE detector, and for each keypoint we extract a descriptor via CNN. N/A Anonymous N/A 256 float32
Brisk + SSS
kp:8000, match:nn		 19-05-14 F 85.8 647.3 93.6 2.31 0.0319 0.0887 0.1437 0.1962 0.2451 — Anonymous We use OpenCV's implementation of brisk detector with the default settings, and for each image there are at most 8K keypoints. For each keypoint, we extract a descriptor via a CNN model. TBA Anonymous N/A 128 float32
D2-Net (single scale)
kp:8000, match:nn		 19-05-07 F 90.1 1398.3 99.0 2.46 0.1201 0.2271 0.3084 0.3732 0.4297 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multiscale)
kp:8000, match:nn		 19-05-07 F 91.7 1757.1 99.5 2.42 0.1115 0.2278 0.3110 0.3786 0.4410 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (single scale, no PT dataset)
kp:8000, match:nn		 19-06-01 F 89.9 1426.3 99.4 2.46 0.1185 0.2385 0.3244 0.3879 0.4425 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multi-scale, no PT dataset)
kp:8000, match:nn		 19-06-05 F 91.2 1608.4 99.5 2.41 0.1038 0.2254 0.3156 0.3833 0.4438 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
DELF
kp:1024, match:nn		 19-05-05 F 75.4 192.4 74.4 2.17 0.0073 0.0340 0.0661 0.1044 0.1389 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:2048, match:nn		 19-05-05 F 79.9 332.0 87.8 2.16 0.0125 0.0466 0.0920 0.1366 0.1807 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:256, match:nn		 19-05-05 F 45.2 45.8 4.3 1.63 0.0000 0.0007 0.0021 0.0028 0.0036 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:512, match:nn		 19-05-05 F 67.8 116.5 41.8 2.17 0.0029 0.0109 0.0221 0.0376 0.0528 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
ELF-256D
kp:512, match:nn		 19-05-07 F 54.8 87.3 14.5 1.97 0.0026 0.0091 0.0133 0.0193 0.0254 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-512D
kp:512, match:nn		 19-05-09 F 65.7 104.9 22.0 2.35 0.0051 0.0128 0.0213 0.0283 0.0359 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-SIFT
kp:512, match:nn		 19-04-26 F 70.5 122.5 38.7 2.45 0.0218 0.0463 0.0665 0.0827 0.0979 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are HOG (as in SIFT). N/A Anonymous N/A 128 uint8
SIFT + GeoDesc
kp:2048, match:nn		 19-05-19 F 86.4 449.6 95.4 2.58 0.1208 0.2030 0.2613 0.3179 0.3589 — Challenge organizers GeoDesc extracted on SIFT keypoints. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:2048, match:nn		 19-05-19 F 86.0 477.3 95.4 2.55 0.1182 0.2001 0.2559 0.3066 0.3520 — Challenge organizers HardNet extracted on SIFT keypoints. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger patches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
HarrisZ/RsGLOH2
kp:8000, match:sGOr2f		 19-05-23 F 92.9 616.1 99.5 2.70 0.2201 0.3324 0.4084 0.4679 0.5146 — Fabio Bellavia and Carlo Colombo HarrisZ keypoints and Root squared (like RootSIFT) sGLOH2 descriptors using sGOr2f* matching strategy. Distance table entries lower than their respective row-wise and column-wise averages are discarded and then matching pairs are computed using the greedy nearest neighbour as in the WISW@CAIP2019 contest. Keypoints and matches are ordered according to their ranks (best ones first). http://cvg.dsi.unifi.it/cvg/index.php?id=research#descriptor bellavia.fabio@gmail.com N/A 256 float32
HesAffNet - HardNet2
kp:8000, match:nn		 19-05-29 F 93.7 1956.7 99.8 2.74 0.2687 0.3719 0.4400 0.4977 0.5402 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian + AffNet(affine shape) + OriNet (orientation), Code and weight from https://github.com/ducha-aiki/affnet, Paper: https://arxiv.org/abs/1711.06704; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset, Paper: https://arxiv.org/pdf/1901.09780.pdf; DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
Hessian - HardNet2
kp:8000, match:nn		 19-05-30 F 91.0 2189.0 99.3 2.74 0.2598 0.3574 0.4140 0.4573 0.4961 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian. Gravity vector orientation is assumed; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset Paper: https://arxiv.org/pdf/1901.09780.pdf, DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
ORB (OpenCV)
kp:8000, match:nn		 19-04-24 F 79.4 885.4 77.7 2.41 0.0379 0.0777 0.1145 0.1500 0.1873 — Challenge organizers ORB, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
Scale-invariant desc. (Log-Polar, lambda=32)
kp:8000, match:nn		 19-06-25 F 95.6 2135.6 99.8 2.66 0.2715 0.3778 0.4451 0.4985 0.5438 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=64)
kp:8000, match:nn		 19-06-24 F 94.6 2171.1 99.8 2.63 0.2515 0.3575 0.4277 0.4835 0.5282 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=96)
kp:8000, match:nn		 19-06-20 F 94.9 2218.2 99.8 2.64 0.2564 0.3650 0.4391 0.4922 0.5377 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SIFT-AID (NN matcher)
kp:8000, match:nn		 19-05-10 F 79.8 580.4 84.3 2.35 0.0425 0.0841 0.1220 0.1612 0.1946 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT-AID (custom matcher)
kp:8000, match:sift-aid		 19-04-29 F/M 78.0 522.7 81.2 2.31 0.0374 0.0808 0.1160 0.1485 0.1837 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT + ContextDesc
kp:8000, match:nn		 19-05-09 F 95.9 1978.9 99.8 2.64 0.2660 0.3681 0.4392 0.4932 0.5405 — Zixin Luo ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 uint8
SIFT-Dense-ContextDesc
kp:8000, match:nn		 19-05-28 F 93.0 1665.6 99.5 2.60 0.2326 0.3430 0.4068 0.4595 0.5065 — Zixin Luo, Jiahui Zhang Dense-ContextDesc is a variant of ContextDesc, where descriptors are densely extracted from full images, instead of image patches, while other settings stay unchanged as original ContextDesc. We find Dense-ContextDesc performs better regarding in particular illumination changes. Dense-ContextDesc is extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. The model will be available on the authors' GitHub page. TBA zluoag@cse.ust.hk N/A TBA
SIFT + ContextDesc + Inlier Classification V2
kp:8000, match:custom		 19-05-28 F/M 97.5 2200.2 99.4 2.79 0.5305 0.6532 0.7134 0.7551 0.7806 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an improved inlier classification and fundamental matrix estimation network based on [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT-GeoDesc-GitHub
kp:8000, match:nn		 19-05-08 F 94.4 1534.7 99.5 2.67 0.2532 0.3562 0.4243 0.4765 0.5245 — Zixin Luo GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features extracted from the code provided by the authors. https://arxiv.org/abs/1807.06294 zluoag@cse.ust.hk N/A 128 float32
SIFT + GeoDesc
kp:8000, match:nn		 19-04-24 F 94.5 1691.3 99.8 2.70 0.2635 0.3651 0.4305 0.4862 0.5288 — Challenge organizers GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:8000, match:nn		 19-04-24 F 96.1 2172.1 100.0 2.68 0.2806 0.3840 0.4511 0.5050 0.5492 — Challenge organizers HardNet extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
SIFT + L2-Net
kp:8000, match:nn		 19-04-24 F 94.1 1850.6 99.8 2.63 0.2415 0.3403 0.4057 0.4598 0.5081 — Challenge organizers L2-Net extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/yuruntian/L2-Net imagematching@uvic.ca N/A 128 float32
SIFT + ContextDesc + Inlier Classification V1
kp:8000, match:custom		 19-05-29 F/M 97.3 2145.2 99.7 2.79 0.4968 0.6233 0.6828 0.7237 0.7565 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an inlier classification and fundamental matrix estimation network using the architecture of [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT (OpenCV)
kp:8000, match:nn		 19-04-24 F 87.0 1011.7 96.6 2.54 0.1358 0.2091 0.2674 0.3167 0.3615 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
SIFT + TFeat
kp:8000, match:nn		 19-04-24 F 92.3 1473.5 99.5 2.58 0.1767 0.2686 0.3347 0.3901 0.4380 — Challenge organizers T-Feat extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/vbalnt/tfeat imagematching@uvic.ca N/A 128 float32
SIFT (OpenCV)
kp:2048, match:nn		 19-05-17 F 77.2 355.9 76.5 2.43 0.0436 0.0817 0.1171 0.1492 0.1800 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
Superpoint (nn matcher)
kp:2048, match:nn		 19-06-07 F 89.3 551.8 98.2 2.77 0.2004 0.3035 0.3649 0.4165 0.4609 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
Superpoint (1:1 matcher)
kp:2048, match:nn1to1		 19-06-07 F 91.0 546.0 98.2 2.86 0.2793 0.3926 0.4616 0.5140 0.5549 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
SuperPoint (default)
kp:2048, match:nn		 19-04-24 F 85.1 350.3 93.0 2.69 0.1697 0.2595 0.3138 0.3585 0.3955 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned (about 1200 on average). Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:1024, match:nn		 19-04-26 F 82.9 302.3 89.0 2.67 0.1578 0.2437 0.2992 0.3403 0.3747 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:2048, match:nn		 19-04-26 F 85.5 503.0 95.2 2.64 0.1462 0.2404 0.3063 0.3578 0.3966 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:256, match:nn		 19-04-26 F 56.7 76.5 28.3 2.10 0.0216 0.0397 0.0505 0.0611 0.0702 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:512, match:nn		 19-04-26 F 78.0 163.2 67.5 2.61 0.1058 0.1604 0.1977 0.2287 0.2527 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:8000, match:nn		 19-04-26 F 89.2 1597.1 98.7 2.54 0.1085 0.2109 0.2856 0.3457 0.3946 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
Scale-invariant desc. (Cartesian, lambda=16)
kp:8000, match:nn		 19-07-29 F 95.8 2303.6 99.9 2.70 0.2463 0.3597 0.4327 0.4821 0.5279 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. (This is a baseline where we use cartesian patches instead.) Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SuperPoint (trained on coco + phototourism training set)
kp:2048, match:nn		 19-05-30 F 87.8 411.3 97.4 2.70 0.1846 0.2857 0.3524 0.4057 0.4491 — Daniel DeTone, Paul Sarlin, Tomasz Malisiewicz, Andrew Rabinovich SuperPoint V1 model trained on COCO homographic warps at VGA resolution, plus pairs from the phototourism training set using the GT poses and depths for correspondence. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork ddetone@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v1)
kp:2048, match:custom		 19-05-30 F/M 92.9 426.6 98.0 2.93 0.3789 0.5105 0.5791 0.6239 0.6568 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set; - keypoints refinement; - better descriptor sampling; - adjusted thresholds. A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v2)
kp:2048, match:custom		 19-05-28 F/M 92.5 514.8 98.5 2.95 0.4132 0.5445 0.6047 0.6475 0.6779 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set, - keypoints refinement, - better descriptor sampling, - adjusted thresholds; A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SURF (OpenCV)
kp:8000, match:nn		 19-04-24 F 80.4 599.7 86.1 2.39 0.0556 0.1045 0.1451 0.1811 0.2148 — Challenge organizers SURF, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SIFT + ContextDesc
kp:8000, match:nn1to1		 19-06-07 F 96.6 2126.4 100.0 2.76 0.3405 0.4583 0.5313 0.5815 0.6252 — Challenge organizers ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with nearest-neighbour search, enforcing cross-match consistency. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc imagematching@uvic.ca N/A TBA
MVS, bag size 10
Method Date Type Ims (%) #Pts SR TL mAP5o mAP10o mAP15o mAP20o mAP25o ATE By Details Link Contact Updated Descriptor size
AKAZE (OpenCV)
kp:8000, match:nn		 19-04-24 F 92.3 3427.4 99.9 3.29 0.3557 0.4639 0.5297 0.5762 0.6130 — Challenge organizers AKAZE, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SphereDesc
kp:8000, match:nn		 19-04-26 F 95.9 3506.0 99.9 3.32 0.3802 0.4957 0.5641 0.6136 0.6536 — Anonymous We use OpenCV's implementation of AKAZE detector, and for each keypoint we extract a descriptor via CNN. N/A Anonymous N/A 256 float32
Brisk + SSS
kp:8000, match:nn		 19-05-14 F 90.3 1900.5 99.8 2.76 0.1190 0.2231 0.2984 0.3588 0.4107 — Anonymous We use OpenCV's implementation of brisk detector with the default settings, and for each image there are at most 8K keypoints. For each keypoint, we extract a descriptor via a CNN model. TBA Anonymous N/A 128 float32
D2-Net (single scale)
kp:8000, match:nn		 19-05-07 F 92.1 3712.5 100.0 3.08 0.2573 0.3867 0.4683 0.5274 0.5741 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multiscale)
kp:8000, match:nn		 19-05-07 F 93.2 4908.6 100.0 2.93 0.2364 0.3763 0.4650 0.5272 0.5776 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (single scale, no PT dataset)
kp:8000, match:nn		 19-06-01 F 92.4 3726.3 100.0 3.08 0.2649 0.4004 0.4782 0.5334 0.5771 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multi-scale, no PT dataset)
kp:8000, match:nn		 19-06-05 F 93.1 4446.3 100.0 2.92 0.2228 0.3619 0.4511 0.5163 0.5672 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
DELF
kp:1024, match:nn		 19-05-05 F 72.1 401.3 96.6 2.34 0.0155 0.0656 0.1214 0.1759 0.2229 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:2048, match:nn		 19-05-05 F 81.5 807.7 99.6 2.33 0.0258 0.0929 0.1647 0.2300 0.2861 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:256, match:nn		 19-05-05 F 30.5 65.3 20.5 1.91 0.0003 0.0019 0.0044 0.0075 0.0103 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:512, match:nn		 19-05-05 F 53.6 170.0 79.0 2.29 0.0046 0.0250 0.0534 0.0817 0.1069 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
ELF-256D
kp:512, match:nn		 19-05-07 F 42.6 130.8 46.8 2.33 0.0192 0.0354 0.0476 0.0586 0.0686 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-512D
kp:512, match:nn		 19-05-09 F 46.6 143.7 61.3 2.55 0.0197 0.0446 0.0634 0.0781 0.0900 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-SIFT
kp:512, match:nn		 19-04-26 F 59.1 210.8 77.5 2.78 0.0837 0.1366 0.1658 0.1890 0.2079 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are HOG (as in SIFT). N/A Anonymous N/A 128 uint8
SIFT + GeoDesc
kp:2048, match:nn		 19-05-19 F 89.8 1221.2 100.0 3.18 0.2894 0.3881 0.4501 0.4996 0.5374 — Challenge organizers GeoDesc extracted on SIFT keypoints. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:2048, match:nn		 19-05-19 F 89.8 1325.4 99.9 3.13 0.2859 0.3818 0.4413 0.4906 0.5285 — Challenge organizers HardNet extracted on SIFT keypoints. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger patches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
HarrisZ/RsGLOH2
kp:8000, match:sGOr2f		 19-05-23 F 95.0 1492.4 100.0 3.41 0.4054 0.5135 0.5792 0.6274 0.6662 — Fabio Bellavia and Carlo Colombo HarrisZ keypoints and Root squared (like RootSIFT) sGLOH2 descriptors using sGOr2f* matching strategy. Distance table entries lower than their respective row-wise and column-wise averages are discarded and then matching pairs are computed using the greedy nearest neighbour as in the WISW@CAIP2019 contest. Keypoints and matches are ordered according to their ranks (best ones first). http://cvg.dsi.unifi.it/cvg/index.php?id=research#descriptor bellavia.fabio@gmail.com N/A 256 float32
HesAffNet - HardNet2
kp:8000, match:nn		 19-05-29 F 96.0 4845.6 100.0 3.55 0.4468 0.5447 0.6021 0.6439 0.6794 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian + AffNet(affine shape) + OriNet (orientation), Code and weight from https://github.com/ducha-aiki/affnet, Paper: https://arxiv.org/abs/1711.06704; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset, Paper: https://arxiv.org/pdf/1901.09780.pdf; DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
Hessian - HardNet2
kp:8000, match:nn		 19-05-30 F 93.8 5326.1 100.0 3.62 0.4358 0.5302 0.5831 0.6210 0.6492 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian. Gravity vector orientation is assumed; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset Paper: https://arxiv.org/pdf/1901.09780.pdf, DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
ORB (OpenCV)
kp:8000, match:nn		 19-04-24 F 79.9 2076.1 98.7 2.86 0.1223 0.1873 0.2356 0.2782 0.3162 — Challenge organizers ORB, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
Scale-invariant desc. (Log-Polar, lambda=32)
kp:8000, match:nn		 19-06-25 F 96.9 5835.6 99.9 3.40 0.4561 0.5610 0.6235 0.6695 0.7032 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=64)
kp:8000, match:nn		 19-06-24 F 97.3 6055.3 100.0 3.39 0.4646 0.5711 0.6309 0.6772 0.7124 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=96)
kp:8000, match:nn		 19-06-20 F 97.5 6132.3 99.7 3.40 0.4750 0.5784 0.6380 0.6810 0.7133 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SIFT-AID (NN matcher)
kp:8000, match:nn		 19-05-10 F 82.2 1693.3 99.9 2.82 0.1595 0.2445 0.2993 0.3459 0.3874 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT-AID (custom matcher)
kp:8000, match:sift-aid		 19-04-29 F/M 79.6 1492.5 99.8 2.73 0.1454 0.2243 0.2778 0.3195 0.3578 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT + ContextDesc
kp:8000, match:nn		 19-05-09 F 97.6 5440.2 99.9 3.35 0.4707 0.5780 0.6369 0.6823 0.7154 — Zixin Luo ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 uint8
SIFT-Dense-ContextDesc
kp:8000, match:nn		 19-05-28 F 95.9 4494.7 100.0 3.36 0.4241 0.5325 0.5955 0.6414 0.6780 — Zixin Luo, Jiahui Zhang Dense-ContextDesc is a variant of ContextDesc, where descriptors are densely extracted from full images, instead of image patches, while other settings stay unchanged as original ContextDesc. We find Dense-ContextDesc performs better regarding in particular illumination changes. Dense-ContextDesc is extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. The model will be available on the authors' GitHub page. TBA zluoag@cse.ust.hk N/A TBA
SIFT + ContextDesc + Inlier Classification V2
kp:8000, match:custom		 19-05-28 F/M 98.2 5753.4 99.9 3.57 0.6926 0.7735 0.8112 0.8365 0.8552 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an improved inlier classification and fundamental matrix estimation network based on [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT-GeoDesc-GitHub
kp:8000, match:nn		 19-05-08 F 97.0 4538.6 99.7 3.43 0.4603 0.5593 0.6158 0.6602 0.6965 — Zixin Luo GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features extracted from the code provided by the authors. https://arxiv.org/abs/1807.06294 zluoag@cse.ust.hk N/A 128 float32
SIFT + GeoDesc
kp:8000, match:nn		 19-04-24 F 96.7 4902.4 99.9 3.49 0.4729 0.5741 0.6312 0.6751 0.7081 — Challenge organizers GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:8000, match:nn		 19-04-24 F 97.7 5954.2 100.0 3.46 0.4768 0.5833 0.6412 0.6865 0.7188 — Challenge organizers HardNet extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
SIFT + L2-Net
kp:8000, match:nn		 19-04-24 F 97.1 5226.7 100.0 3.36 0.4169 0.5218 0.5862 0.6348 0.6738 — Challenge organizers L2-Net extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/yuruntian/L2-Net imagematching@uvic.ca N/A 128 float32
SIFT + ContextDesc + Inlier Classification V1
kp:8000, match:custom		 19-05-29 F/M 97.8 5654.6 100.0 3.56 0.6630 0.7521 0.7933 0.8216 0.8422 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an inlier classification and fundamental matrix estimation network using the architecture of [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT (OpenCV)
kp:8000, match:nn		 19-04-24 F 91.0 3231.3 100.0 3.20 0.3372 0.4350 0.4961 0.5408 0.5769 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
SIFT + TFeat
kp:8000, match:nn		 19-04-24 F 96.2 4513.6 99.9 3.27 0.3855 0.4961 0.5620 0.6125 0.6535 — Challenge organizers T-Feat extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/vbalnt/tfeat imagematching@uvic.ca N/A 128 float32
SIFT (OpenCV)
kp:2048, match:nn		 19-05-17 F 75.0 862.7 99.1 2.91 0.1481 0.2135 0.2591 0.2959 0.3275 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
Superpoint (nn matcher)
kp:2048, match:nn		 19-06-07 F 93.1 1391.0 99.8 3.64 0.3915 0.5034 0.5674 0.6124 0.6491 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
Superpoint (1:1 matcher)
kp:2048, match:nn1to1		 19-06-07 F 93.9 1368.9 100.0 3.79 0.4608 0.5709 0.6303 0.6724 0.7042 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
SuperPoint (default)
kp:2048, match:nn		 19-04-24 F 87.4 851.9 99.9 3.53 0.3440 0.4490 0.5066 0.5486 0.5818 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned (about 1200 on average). Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:1024, match:nn		 19-04-26 F 85.0 710.3 99.9 3.50 0.3409 0.4379 0.4946 0.5349 0.5643 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:2048, match:nn		 19-04-26 F 88.6 1242.8 100.0 3.46 0.3292 0.4353 0.4977 0.5417 0.5752 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:256, match:nn		 19-04-26 F 50.5 135.1 54.0 2.64 0.0793 0.1144 0.1352 0.1486 0.1596 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:512, match:nn		 19-04-26 F 71.9 327.1 94.6 3.24 0.2366 0.3165 0.3608 0.3903 0.4135 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:8000, match:nn		 19-04-26 F 92.0 4162.5 100.0 3.26 0.2501 0.3775 0.4557 0.5122 0.5573 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
Scale-invariant desc. (Cartesian, lambda=16)
kp:8000, match:nn		 19-07-29 F 98.1 6295.4 99.9 3.43 0.4245 0.5321 0.5971 0.6453 0.6827 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. (This is a baseline where we use cartesian patches instead.) Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SuperPoint (trained on coco + phototourism training set)
kp:2048, match:nn		 19-05-30 F 92.1 1061.1 100.0 3.60 0.3959 0.5095 0.5732 0.6210 0.6569 — Daniel DeTone, Paul Sarlin, Tomasz Malisiewicz, Andrew Rabinovich SuperPoint V1 model trained on COCO homographic warps at VGA resolution, plus pairs from the phototourism training set using the GT poses and depths for correspondence. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork ddetone@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v1)
kp:2048, match:custom		 19-05-30 F/M 94.8 1043.4 100.0 3.99 0.5651 0.6701 0.7229 0.7581 0.7830 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set; - keypoints refinement; - better descriptor sampling; - adjusted thresholds. A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v2)
kp:2048, match:custom		 19-05-28 F/M 95.1 1240.8 100.0 4.01 0.5774 0.6846 0.7348 0.7685 0.7925 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set, - keypoints refinement, - better descriptor sampling, - adjusted thresholds; A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SURF (OpenCV)
kp:8000, match:nn		 19-04-24 F 85.0 1790.2 99.9 2.91 0.1962 0.2835 0.3396 0.3866 0.4253 — Challenge organizers SURF, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SIFT + ContextDesc
kp:8000, match:nn1to1		 19-06-07 F 97.5 5894.8 99.9 3.46 0.5135 0.6222 0.6796 0.7206 0.7522 — Challenge organizers ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with nearest-neighbour search, enforcing cross-match consistency. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc imagematching@uvic.ca N/A TBA
MVS, bag size 25
Method Date Type Ims (%) #Pts SR TL mAP5o mAP10o mAP15o mAP20o mAP25o ATE By Details Link Contact Updated Descriptor size
AKAZE (OpenCV)
kp:8000, match:nn		 19-04-24 F 96.9 12204.1 99.9 4.86 0.6385 0.7371 0.7845 0.8156 0.8382 — Challenge organizers AKAZE, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SphereDesc
kp:8000, match:nn		 19-04-26 F 97.8 10035.3 100.0 4.75 0.6077 0.7228 0.7800 0.8159 0.8414 — Anonymous We use OpenCV's implementation of AKAZE detector, and for each keypoint we extract a descriptor via CNN. N/A Anonymous N/A 256 float32
Brisk + SSS
kp:8000, match:nn		 19-05-14 F 95.5 7221.3 100.0 3.83 0.3383 0.4770 0.5552 0.6111 0.6545 — Anonymous We use OpenCV's implementation of brisk detector with the default settings, and for each image there are at most 8K keypoints. For each keypoint, we extract a descriptor via a CNN model. TBA Anonymous N/A 128 float32
D2-Net (single scale)
kp:8000, match:nn		 19-05-07 F 93.4 11248.3 100.0 4.46 0.4339 0.5825 0.6538 0.6982 0.7290 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multiscale)
kp:8000, match:nn		 19-05-07 F 93.6 14569.4 100.0 4.16 0.3954 0.5507 0.6284 0.6768 0.7105 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Trained on sequences overlapping with our test set: see 'no PT' for eligible results (this entry is provided only for reference). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (single scale, no PT dataset)
kp:8000, match:nn		 19-06-01 F 93.5 11241.5 100.0 4.46 0.4305 0.5784 0.6514 0.6971 0.7292 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Single-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
D2-Net (multi-scale, no PT dataset)
kp:8000, match:nn		 19-06-05 F 93.6 13525.8 100.0 4.10 0.3872 0.5469 0.6263 0.6749 0.7089 — Challenge organizers D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. Multi-scale features, brute-force nearest neighbour matching. Re-trained removing conflicting sequences (models: d2_tf_no_phototourism.pth). Paper: https://dsmn.ml/files/d2-net/d2-net.pdf https://github.com/mihaidusmanu/d2-net imagematching@uvic.ca N/A 512 float32
DELF
kp:1024, match:nn		 19-05-05 F 85.3 1810.3 100.0 2.85 0.0577 0.1710 0.2682 0.3425 0.3984 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:2048, match:nn		 19-05-05 F 90.3 3748.5 100.0 2.84 0.0863 0.2245 0.3300 0.4077 0.4649 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:256, match:nn		 19-05-05 F 26.2 121.5 63.5 2.27 0.0022 0.0107 0.0219 0.0336 0.0437 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
DELF
kp:512, match:nn		 19-05-05 F 66.0 681.4 94.5 2.71 0.0206 0.0812 0.1465 0.2013 0.2456 — Challenge organizers DELF features for object retrieval, trained on the Google Landmarks dataset. Paper: https://arxiv.org/abs/1612.06321 https://github.com/tensorflow/models/tree/master/research/delf imagematching@uvic.ca N/A 40 float32
ELF-256D
kp:512, match:nn		 19-05-07 F 49.6 340.4 89.6 3.32 0.1175 0.1690 0.1960 0.2164 0.2325 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-512D
kp:512, match:nn		 19-05-09 F 56.5 373.9 94.1 3.54 0.1351 0.2074 0.2455 0.2714 0.2912 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are computed with the interpolation of the VGG pool3 feature map on the detected keypoints. TBA Anonymous N/A 256 float32
ELF-SIFT
kp:512, match:nn		 19-04-26 F 69.5 662.0 99.9 3.89 0.2922 0.3696 0.4087 0.4356 0.4549 — Anonymous ELF detector: Keypoints are local maxima of a saliency map generated by the gradient of a feature map with respect to the image of a pre-trained CNN. Descriptors are HOG (as in SIFT). N/A Anonymous N/A 128 uint8
SIFT + GeoDesc
kp:2048, match:nn		 19-05-19 F 95.2 3924.3 100.0 4.46 0.5081 0.6074 0.6626 0.7023 0.7336 — Challenge organizers GeoDesc extracted on SIFT keypoints. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:2048, match:nn		 19-05-19 F 95.8 4446.8 100.0 4.39 0.5103 0.6106 0.6677 0.7073 0.7380 — Challenge organizers HardNet extracted on SIFT keypoints. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger patches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
HarrisZ/RsGLOH2
kp:8000, match:sGOr2f		 19-05-23 F 98.4 4448.9 100.0 4.79 0.6474 0.7517 0.8013 0.8325 0.8555 — Fabio Bellavia and Carlo Colombo HarrisZ keypoints and Root squared (like RootSIFT) sGLOH2 descriptors using sGOr2f* matching strategy. Distance table entries lower than their respective row-wise and column-wise averages are discarded and then matching pairs are computed using the greedy nearest neighbour as in the WISW@CAIP2019 contest. Keypoints and matches are ordered according to their ranks (best ones first). http://cvg.dsi.unifi.it/cvg/index.php?id=research#descriptor bellavia.fabio@gmail.com N/A 256 float32
HesAffNet - HardNet2
kp:8000, match:nn		 19-05-29 F 97.7 13789.1 100.0 5.13 0.6736 0.7711 0.8154 0.8442 0.8644 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian + AffNet(affine shape) + OriNet (orientation), Code and weight from https://github.com/ducha-aiki/affnet, Paper: https://arxiv.org/abs/1711.06704; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset, Paper: https://arxiv.org/pdf/1901.09780.pdf; DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
Hessian - HardNet2
kp:8000, match:nn		 19-05-30 F 95.5 14054.5 100.0 5.26 0.6298 0.7190 0.7583 0.7817 0.7978 — Milan Pultar, Dmytro Mishkin, Jiří Matas Detector: Hessian. Gravity vector orientation is assumed; Descriptor: HardNet trained on AMOS + mix of other datasets, similar to Code for train: https://github.com/pultarmi/HardNet_MultiDataset Paper: https://arxiv.org/pdf/1901.09780.pdf, DeepNets were plugged into MODS framework, but without view synthesis and matching https://github.com/ducha-aiki/mods-light-zmq; Number of max.keypoints to detect: 8k, detection done on 2x upsampled images. TBA ducha.aiki@gmail.com N/A 128 uint8
ORB (OpenCV)
kp:8000, match:nn		 19-04-24 F 91.3 8300.6 99.9 3.75 0.3708 0.4608 0.5174 0.5603 0.5951 — Challenge organizers ORB, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
Scale-invariant desc. (Log-Polar, lambda=32)
kp:8000, match:nn		 19-06-25 F 98.2 17456.1 100.0 4.91 0.7009 0.7902 0.8312 0.8573 0.8764 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=64)
kp:8000, match:nn		 19-06-24 F 98.3 17795.1 100.0 4.89 0.7073 0.8003 0.8411 0.8663 0.8841 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
Scale-invariant desc. (Log-Polar, lambda=96)
kp:8000, match:nn		 19-06-20 F 98.1 17834.9 100.0 4.88 0.7028 0.7954 0.8363 0.8612 0.8783 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SIFT-AID (NN matcher)
kp:8000, match:nn		 19-05-10 F 92.9 6882.4 100.0 4.02 0.4169 0.5321 0.5940 0.6389 0.6739 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT-AID (custom matcher)
kp:8000, match:sift-aid		 19-04-29 F/M 92.7 6640.7 100.0 3.88 0.3907 0.5078 0.5717 0.6174 0.6533 — Mariano Rodríguez, Gabriele Facciolo, Rafael Grompone Von Gioi, Pablo Musé, Jean-Michel Morel, Julie Delon We extract the keypoints using OpenCV's implementation of SIFT. The AID descriptors are computed with a CNN from patches extracted at each keypoint location, the result is a binary descriptor of 6272 bits. The matching is computed as the Hamming distance between the descriptors, with the decision threshold set at 4000. Preprint: https://hal.archives-ouvertes.fr/hal-02016010. Code: https://github.com/rdguez-mariano/sift-aid https://hal.archives-ouvertes.fr/hal-02016010 facciolo@cmla.ens-cachan.fr N/A 6272 bits
SIFT + ContextDesc
kp:8000, match:nn		 19-05-09 F 98.1 15670.4 100.0 4.78 0.6910 0.7879 0.8295 0.8549 0.8727 — Zixin Luo ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 uint8
SIFT-Dense-ContextDesc
kp:8000, match:nn		 19-05-28 F 97.6 13254.7 100.0 4.94 0.6745 0.7753 0.8197 0.8463 0.8648 — Zixin Luo, Jiahui Zhang Dense-ContextDesc is a variant of ContextDesc, where descriptors are densely extracted from full images, instead of image patches, while other settings stay unchanged as original ContextDesc. We find Dense-ContextDesc performs better regarding in particular illumination changes. Dense-ContextDesc is extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features are quantized to uint8 and extracted from the code provided by the authors. The model will be available on the authors' GitHub page. TBA zluoag@cse.ust.hk N/A TBA
SIFT + ContextDesc + Inlier Classification V2
kp:8000, match:custom		 19-05-28 F/M 98.7 15528.9 100.0 5.06 0.8172 0.8889 0.9144 0.9281 0.9369 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an improved inlier classification and fundamental matrix estimation network based on [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT-GeoDesc-GitHub
kp:8000, match:nn		 19-05-08 F 98.3 13975.1 100.0 4.95 0.7162 0.8077 0.8460 0.8697 0.8861 — Zixin Luo GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. Features extracted from the code provided by the authors. https://arxiv.org/abs/1807.06294 zluoag@cse.ust.hk N/A 128 float32
SIFT + GeoDesc
kp:8000, match:nn		 19-04-24 F 98.2 14861.6 99.8 5.09 0.7163 0.8006 0.8391 0.8634 0.8804 — Challenge organizers GeoDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://github.com/lzx551402/geodesc imagematching@uvic.ca N/A 128 float32
SIFT + HardNet
kp:8000, match:nn		 19-04-24 F 97.9 17011.3 100.0 5.00 0.6875 0.7833 0.8263 0.8535 0.8720 — Challenge organizers HardNet extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/DagnyT/hardnet imagematching@uvic.ca N/A 128 float32
SIFT + L2-Net
kp:8000, match:nn		 19-04-24 F 98.0 16304.6 99.9 4.84 0.6672 0.7666 0.8120 0.8409 0.8613 — Challenge organizers L2-Net extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/yuruntian/L2-Net imagematching@uvic.ca N/A 128 float32
SIFT + ContextDesc + Inlier Classification V1
kp:8000, match:custom		 19-05-29 F/M 98.7 15390.7 100.0 5.05 0.8063 0.8783 0.9053 0.9208 0.9311 — Dawei Sun, Zixin Luo, Jiahui Zhang We use the SIFT detector and ContextDesc descriptor, and then we train an inlier classification and fundamental matrix estimation network using the architecture of [Yi et al. CVPR2018] (https://arxiv.org/pdf/1711.05971.pdf). https://github.com/lzx551402/contextdesc zluoag@cse.ust.hk N/A 128 float32
SIFT (OpenCV)
kp:8000, match:nn		 19-04-24 F 96.5 12489.7 99.8 4.61 0.6326 0.7237 0.7679 0.7976 0.8201 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
SIFT + TFeat
kp:8000, match:nn		 19-04-24 F 97.3 14948.2 99.9 4.70 0.6368 0.7349 0.7818 0.8120 0.8333 — Challenge organizers T-Feat extracted on SIFT keypoints. Number of keypoints: 8000 per image. Models trained on the Liberty sequence of the Brown dataset. We use slightly larger paches than specified for SIFT (scale multiplying factor 16/12). Feature matching with brute-force nearest-neighbour search. https://github.com/vbalnt/tfeat imagematching@uvic.ca N/A 128 float32
SIFT (OpenCV)
kp:2048, match:nn		 19-05-17 F 89.3 3374.9 100.0 4.11 0.4023 0.4912 0.5408 0.5766 0.6053 — Challenge organizers SIFT, as implemented in OpenCV. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A 128 float32
Superpoint (nn matcher)
kp:2048, match:nn		 19-06-07 F 97.4 4048.4 100.0 5.43 0.6195 0.7274 0.7821 0.8167 0.8405 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
Superpoint (1:1 matcher)
kp:2048, match:nn1to1		 19-06-07 F 97.6 3880.2 100.0 5.60 0.6605 0.7619 0.8114 0.8414 0.8625 — Challenge organizers SuperPoint features from the submission 'SuperPoint + Custom Matcher (v2)' with a brute-force 1:1 matcher instead of the custom matcher. For reference. TBA imagematching@uvic.ca N/A 256 float32
SuperPoint (default)
kp:2048, match:nn		 19-04-24 F 92.3 2528.6 100.0 5.42 0.5717 0.6706 0.7169 0.7458 0.7659 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned (about 1200 on average). Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:1024, match:nn		 19-04-26 F 92.3 2041.9 100.0 5.36 0.5732 0.6699 0.7154 0.7438 0.7631 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:2048, match:nn		 19-04-26 F 93.2 3797.1 100.0 5.19 0.5798 0.6805 0.7271 0.7558 0.7752 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:256, match:nn		 19-04-26 F 58.5 340.4 81.3 4.11 0.2275 0.2796 0.3074 0.3264 0.3405 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:512, match:nn		 19-04-26 F 82.8 937.5 100.0 4.93 0.4612 0.5456 0.5885 0.6155 0.6349 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
SuperPoint
kp:8000, match:nn		 19-04-26 F 93.8 13117.2 100.0 4.83 0.4604 0.6008 0.6686 0.7107 0.7396 — Challenge organizers SuperPoint features. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We lower the default detection threshold to take the number of features indicated in the label. Feature matching done with brute-force nearest-neighbour search. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork imagematching@uvic.ca N/A 256 float32
Scale-invariant desc. (Cartesian, lambda=16)
kp:8000, match:nn		 19-07-29 F 97.6 18349.5 100.0 4.87 0.6240 0.7293 0.7811 0.8128 0.8361 — Patrick Ebel We compute scale-invariant descriptors with a log-polar transformation of the patch. Keypoints are DoG, with a scaling factor of lambda/12 over its chosen scale. (This is a baseline where we use cartesian patches instead.) Reference: 'Beyond Cartesian Representations for Local Descriptors', Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls, ICCV 2019. TBA patrick.ebel@epfl.ch N/A 128 float32
SuperPoint (trained on coco + phototourism training set)
kp:2048, match:nn		 19-05-30 F 97.3 3130.6 100.0 5.50 0.6528 0.7599 0.8084 0.8380 0.8581 — Daniel DeTone, Paul Sarlin, Tomasz Malisiewicz, Andrew Rabinovich SuperPoint V1 model trained on COCO homographic warps at VGA resolution, plus pairs from the phototourism training set using the GT poses and depths for correspondence. If necessary, we downsample the images so that the largest dimension is at most 1024 pixels. We extract features with the default parameters and use however many are returned. https://github.com/MagicLeapResearch/SuperPointPretrainedNetwork ddetone@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v1)
kp:2048, match:custom		 19-05-30 F/M 98.2 2813.7 100.0 6.06 0.7527 0.8355 0.8718 0.8923 0.9062 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set; - keypoints refinement; - better descriptor sampling; - adjusted thresholds. A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SuperPoint + Custom Matcher (v2)
kp:2048, match:custom		 19-05-28 F/M 98.2 3358.5 100.0 6.13 0.7609 0.8430 0.8765 0.8952 0.9083 — Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich Features are extracted by a modified SuperPoint: - retrained on the phototourism training set, - keypoints refinement, - better descriptor sampling, - adjusted thresholds; A custom matcher estimates the image rotation and rejects outlier matches using the correspondence classifier introduced in 'Learning To Find Good Correspondences' (Yi et al., 2018). TBA pesarlin@magicleap.com N/A 256 float32
SURF (OpenCV)
kp:8000, match:nn		 19-04-24 F 94.8 8611.8 99.9 3.98 0.4916 0.5971 0.6547 0.6956 0.7271 — Challenge organizers SURF, as implemented in OpenCV. Number of keypoints: 8000 per image. Feature matching with brute-force nearest-neighbour search. https://opencv.org imagematching@uvic.ca N/A TBA
SIFT + ContextDesc
kp:8000, match:nn1to1		 19-06-07 F 98.3 16867.5 100.0 4.85 0.7012 0.8004 0.8415 0.8664 0.8837 — Challenge organizers ContextDesc extracted on SIFT keypoints. Number of keypoints: 8000 per image. Feature matching with nearest-neighbour search, enforcing cross-match consistency. Features are quantized to uint8 and extracted from the code provided by the authors. https://github.com/lzx551402/contextdesc imagematching@uvic.ca N/A TBA
< prev | breakdown | next >

home