Semantic Correspondence on PF-WILLOW
91.3PCK@0.1 (bbox)MATCHA
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| MATCHASup.=✓, Resolution=512x5122025.01 | 91.3 | 70.2 | 97 | — | — | — | — | |
| MATCHASingle Desc=true, Corres. Sup.=GM+SM2025.01 | 91.3 | 70.2 | 97 | 86.2 | — | — | — | |
| MATCHA-LightSingle Desc=false, Corres. Sup.=GM+SM2025.01 | 90.6 | 69 | 96.2 | 85.3 | — | — | — | |
| MATCHA-LightSup.=✓, Resolution=512x5122025.01 | 90.1 | 69 | 96.2 | — | — | — | — | |
| DIFT (ours)Sup.=X, Resolution=512x5122025.01 | 85.1 | 55.7 | 92.9 | — | — | — | — | |
| DIFTSingle Desc=false, Corres. Sup.=X2025.01 | 85.1 | 55.7 | 92.9 | 77.9 | — | — | — | |
| USCSup.=X, Resolution=512x5122025.01 | 84.3 | 53 | — | — | — | — | — | |
| DIFT.Uni + DINOv2Single Desc=false, Corres. Sup.=X2025.01 | 82.9 | 58.7 | 90.7 | 77.4 | — | — | — | |
| VATF.T.=true, Feat.=true, Data Aug.=true, Cost Aggregation=Transformer2022.07 | 81.6 | 52.8 | 91.4 | — | — | — | — | |
| SD4Match-SingleTraining/Tuning Dataset=SPair-71k2023.10 | 81.3 | 55.5 | 91.2 | — | — | — | — | |
| DIFTsdSupervision=(c) no supervision2023.06 | 81.2 | 58.1 | — | — | — | — | — | |
| DIFTSup.=X, Dataset-specific models=true, Resolution=512x5122025.01 | 81.2 | 58.1 | — | — | — | — | — | |
| LPMFlowBackbone=ViT-B/16, Resolution=2562025.05 | 81 | — | — | — | — | — | — | |
| SD4Match-CPMTraining/Tuning Dataset=SPair-71k2023.10 | 80.9 | 56.7 | 91.6 | — | — | — | — | |
| SD4MatchSup.=✓, Dataset-specific models=true, Image pairs as inputs=true, Resolution=512x5122025.01 | 80.9 | 56.7 | 91.6 | — | — | — | — | |
| SD4Match-CPMTraining/Tuning Dataset=PF-Pascal2023.10 | 80.4 | 52.1 | 91.2 | — | — | — | — | |
| SD4MatchBackbone=SD+DINOv2, Resolution=7682025.05 | 80.4 | 52.1 | 91.2 | — | — | — | — | |
| DHPFSupervision=weak, Signal=image-level labels, Backbone=ResNet-101, Inference time=56ms2020.07 | 80.2 | 50.2 | 91.1 | — | — | — | — | |
| SCorrSANSupervision=(a) fully supervised2023.06 | 80 | 54.1 | — | — | — | — | — | |
| SCorrSANSup.=✓, Dataset-specific models=true, Image pairs as inputs=true, Resolution=512x5122025.01 | 80 | 54.1 | 89.8 | — | — | — | — | |
| CHMFeat.-level=Single, FT. feat.=false, Aggregation=6D Conv.2021.06 | 79.4 | 52.7 | 87.5 | — | — | — | — | |
| CHMF.T.=false, Feat.=true, Data Aug.=false, Cost Aggregation=4D Conv.2022.07 | 79.4 | 52.7 | 87.5 | — | — | — | — | |
| CHMSupervision=(a) fully supervised2023.06 | 79.4 | 52.7 | — | — | — | — | — | |
| CATsFeat.-level=Multi, FT. feat.=true, Aggregation=Transformer2021.06 | 79.2 | 50.3 | 90.3 | — | — | — | — | |
| CATSF.T.=true, Feat.=true, Data Aug.=true, Cost Aggregation=Transformer2022.07 | 79.2 | 50.3 | 90.3 | — | — | — | — | |
| CATSSupervision=(a) fully supervised2023.06 | 79.2 | 50.3 | — | — | — | — | — | |
| VATF.T.=true, Feat.=true, Data Aug.=false, Cost Aggregation=Transformer2022.07 | 78.8 | 46 | 91.3 | — | — | — | — | |
| DHPFSupervision=weak, Signal=image-level labels, Backbone=ResNet-101, Inference time=33ms2020.07 | 78.7 | 51.8 | 89.6 | — | — | — | — | |
| GSFFeat.-level=Multi, FT. feat.=false, Aggregation=2D Conv.2021.06 | 78.7 | 49.1 | 90.2 | — | — | — | — | |
| GSFSupervision=(b) weakly supervised2023.06 | 78.7 | 49.1 | — | — | — | — | — | |
| SD4Match-SingleTraining/Tuning Dataset=PF-Pascal2023.10 | 77.8 | 50.7 | 89.5 | — | — | — | — | |
| DHPFSupervision=strong, Signal=src & trg keypoint matches, Backbone=ResNet-101, mu=0.4, Inference time=58ms2020.07 | 77.6 | 49.5 | 89.1 | — | — | — | — | |
| DHPFFeat.-level=Multi, FT. feat.=false, Aggregation=RHM2021.06 | 77.6 | 49.5 | 89.1 | — | — | — | — | |
| DHPFF.T.=false, Feat.=true, Data Aug.=false, Cost Aggregation=RHM2022.07 | 77.6 | 49.5 | 89.1 | — | — | — | — | |
| DHPFSupervision=(a) fully supervised2023.06 | 77.6 | 49.5 | — | — | — | — | — | |
| HPFBackbone=ResNet-101-FCN, Supervision=keypoints (validation only)2019.08 | 76.3 | 48.6 | 88.2 | — | — | — | — | |
| SCOTFeat.-level=Multi, Aggregation=OT-RHM2021.06 | 76 | 47.8 | 87.1 | — | — | — | — | |
| SCOTF.T.=false, Feat.=true, Data Aug.=false, Cost Aggregation=OT-RHM2022.07 | 76 | 47.8 | 87.1 | — | — | — | — | |
| TransforMatcherSupervision=(a) fully supervised2023.06 | 76 | — | — | — | — | — | — | |
| PWarpCSupervision=(b) weakly supervised2023.06 | 75.9 | 45 | — | — | — | — | — | |
| PWarpC-NCNetTraining/Tuning Dataset=PF-Pascal2023.10 | 75.9 | 45 | 87.9 | — | — | — | — | |
| DHPFSupervision=weak, Signal=image-level labels, Backbone=ResNet-50, Inference time=31ms2020.07 | 75.7 | 48.7 | 87.3 | — | — | — | — | |
| CATs†Feat.-level=Multi, FT. feat.=false, Aggregation=Transformer2021.06 | 75.6 | 46.6 | 87.5 | — | — | — | — | |
| CATSF.T.=true, Feat.=true, Data Aug.=false, Cost Aggregation=Transformer2022.07 | 75.6 | 46.6 | 87.5 | — | — | — | — | |
| PMDSupervision=(a) fully supervised2023.06 | 75.6 | — | — | — | — | — | — | |
| DINOv2Sup.=X, Resolution=512x5122025.01 | 75.4 | 43.8 | 86.1 | — | — | — | — | |
| DINOv2Single Desc=true, Corres. Sup.=X2025.01 | 75.4 | 43.8 | 86.1 | 68.4 | — | — | — | |
| WarpCSupervision=(b) weakly supervised2023.06 | 75.1 | 49 | — | — | — | — | — | |
| DHPFSupervision=strong, Signal=src & trg keypoint matches, Backbone=ResNet-101, Inference time=34ms2020.07 | 74.8 | 47.9 | 86.7 | — | — | — | — | |
| HPFBackbone=ResNet-101, Supervision=keypoints (validation only)2019.08 | 74.4 | 45.9 | 85.6 | — | — | — | — | |
| HPFSupervision=strong, Signal=src & trg keypoint matches, Backbone=ResNet-101, Inference time=63ms2020.07 | 74.4 | 45.9 | 85.6 | — | — | — | — | |
| HPFFeat.-level=Multi, Aggregation=RHM2021.06 | 74.4 | 45.9 | 85.6 | — | — | — | — | |
| Ours(DINOv2+ResNet)Backbone=DINOv2, Resolution=840, Additional component=ResNet bottleneck2025.05 | 74.1 | 46.6 | 87.1 | — | — | — | — | |
| SF-NetSupervision=weak, Signal=bbox, Backbone=ResNet-101, Inference time=51ms2020.07 | 74 | 46.3 | 84.2 | — | — | — | — | |
| SFNetFeat.-level=Multi, FT. feat.=false, Aggregation=-2021.06 | 74 | 46.3 | 84.2 | — | — | — | — | |
| SFNetTraining/Tuning Dataset=PF-Pascal2023.10 | 74 | 46.3 | 84.2 | — | — | — | — | |
| DCC-NetSupervision=weak, Signal=image-level labels, Backbone=ResNet-101, Inference time=>261ms2020.07 | 73.8 | 43.6 | 86.5 | — | — | — | — | |
| DCC-NetFeat.-level=Single, FT. feat.=false, Aggregation=4D Conv.2021.06 | 73.8 | 43.6 | 86.5 | — | — | — | — | |
| Ours(DINOv2)Backbone=DINOv2, Resolution=8402025.05 | 73.7 | 48.8 | 85.8 | — | — | — | — | |
| HPFBackbone=ResNet-101, Hyperparameter k=3, Supervision=keypoints (validation only, small set)2019.08 | 73.4 | 45.1 | 85.4 | — | — | — | — | |
| DHPFSupervision=strong, Signal=src & trg keypoint matches, Backbone=ResNet-50, Inference time=30ms2020.07 | 73.3 | 45.8 | 86.6 | — | — | — | — | |
| HPFBackbone=ResNet-101, Hyperparameter k=2, Supervision=keypoints (validation only, small set)2019.08 | 73.1 | 44.7 | 85.4 | — | — | — | — | |
| NCNetTraining/Tuning Dataset=PF-Pascal2023.10 | 72.7 | 44 | 85.4 | — | — | — | — | |
| NN-CycBackbone=ResNet-1012019.08 | 72.5 | 40.5 | 86.9 | — | — | — | — | |
| HPFBackbone=ResNet-101, Hyperparameter k=1, Supervision=keypoints (validation only, small set)2019.08 | 72.5 | 44.5 | 84.8 | — | — | — | — | |
| HPFBackbone=ResNet-50, Supervision=keypoints (validation only)2019.08 | 72.4 | 46.5 | 84.7 | — | — | — | — | |
| HPFSupervision=strong, Signal=src & trg keypoint matches, Backbone=ResNet-50, Inference time=34ms2020.07 | 72.4 | 46.5 | 84.7 | — | — | — | — | |
| SD+DINOTraining/Tuning Dataset=PF-Pascal (default/zero-shot)2023.10 | 72 | 44.8 | 85 | — | — | — | — | |
| RTNsBackbone=ResNet-1012019.08 | 71.9 | 41.3 | 86.2 | — | — | — | — | |
| RTNSupervision=weak, Signal=bbox, Backbone=ResNet-101, Inference time=376ms2020.07 | 71.9 | 41.3 | 86.2 | — | — | — | — | |
| RTNsFeat.-level=Single, FT. feat.=false, Aggregation=-2021.06 | 71.9 | 41.3 | 86.2 | — | — | — | — | |
| SimSC-iBOTTraining/Tuning Dataset=PF-Pascal2023.10 | 71.4 | 44.9 | 84.5 | — | — | — | — | |
| SimSC-iBOTBackbone=iBOT, Resolution=2562025.05 | 71.4 | 44.9 | 84.5 | — | — | — | — | |
| DHPFTraining/Tuning Dataset=PF-Pascal2023.10 | 70.6 | 44.8 | 83.2 | — | — | — | — | |
| SCNetBackbone=VGG-16, Supervision=keypoints2019.08 | 70.4 | 38.6 | 85.3 | — | — | — | — | |
| SCNetSupervision=strong, Signal=src & trg keypoint matches, Backbone=VGG-16, Inference time=1000ms2020.07 | 70.4 | 38.6 | 85.3 | — | — | — | — | |
| SCNetSupervision=(a) fully supervised2023.06 | 70.4 | 38.6 | — | — | — | — | — | |
| WeakalignBackbone=ResNet-101, Supervision=image labels2019.08 | 70.2 | 37 | 79.9 | — | — | — | — | |
| WeakalignSupervision=weak, Signal=bbox, Backbone=ResNet-101, Inference time=41ms2020.07 | 70.2 | 37 | 79.9 | — | — | — | — | |
| WeakAlignFeat.-level=Single, FT. feat.=false, Aggregation=-2021.06 | 70.2 | 37 | 79.9 | — | — | — | — | |
| CATSTraining/Tuning Dataset=PF-Pascal2023.10 | 69.5 | 40.9 | 83.2 | — | — | — | — | |
| CNNGeoBackbone=ResNet-101, Supervision=synthetic warp2019.08 | 69.2 | 36.9 | 77.8 | — | — | — | — | |
| CNNGeoSupervision=self, Signal=synthetic pairs, Backbone=ResNet-101, Inference time=40ms2020.07 | 69.2 | 36.9 | 77.8 | — | — | — | — | |
| CNNGeoFeat.-level=Single, FT. feat.=false, Aggregation=-2021.06 | 69.2 | 36.9 | 77.8 | — | — | — | — | |
| A2NetBackbone=ResNet-101, Supervision=self-supervised2019.08 | 68.8 | 36.3 | 84.4 | — | — | — | — | |
| A2NetSupervision=self, Signal=synthetic pairs, Backbone=ResNet-101, Inference time=53ms2020.07 | 68.8 | 36.3 | 84.4 | — | — | — | — | |
| A2NetFeat.-level=Single, FT. feat.=false, Aggregation=-2021.06 | 68.8 | 36.3 | 84.4 | — | — | — | — | |
| DIFTTraining/Tuning Dataset=PF-Pascal (default/zero-shot)2023.10 | 68 | 44.8 | 79.8 | — | — | — | — | |
| NC-NetBackbone=ResNet-101, Supervision=weakly-supervised2019.08 | 67 | 33.8 | 83.7 | — | — | — | — | |
| NC-NetSupervision=weak, Signal=image-level labels, Backbone=ResNet-101, Inference time=261ms2020.07 | 67 | 33.8 | 83.7 | — | — | — | — | |
| NC-NetFeat.-level=Single, FT. feat.=true, Aggregation=4D Conv.2021.06 | 67 | 33.8 | 83.7 | — | — | — | — | |
| NC-NetF.T.=true, Feat.=true, Data Aug.=false, Cost Aggregation=4D Conv.2022.07 | 67 | 33.8 | 83.7 | — | — | — | — | |
| DIFTadmSupervision=(c) no supervision2023.06 | 67 | 46.9 | — | — | — | — | — | |
| Trans.MatTraining/Tuning Dataset=PF-Pascal2023.10 | 65.3 | — | — | — | — | — | — | |
| DINOv1Training/Tuning Dataset=PF-Pascal (default/zero-shot)2023.10 | 64.8 | 39.7 | 77 | — | — | — | — | |
| DINOv2Training/Tuning Dataset=PF-Pascal (default/zero-shot)2023.10 | 63.4 | 37.3 | 73.1 | — | — | — | — | |
| HPFBackbone=ResNet-101, Layer selection=random, Supervision=keypoints (validation only, small set)2019.08 | 62.4 | 32.8 | 78.2 | — | — | — | — | |
| Ours(DINOv2+NC)Backbone=DINOv2, Resolution=840, Additional component=NC cost aggregator2025.05 | 62.3 | 39.8 | 74.9 | — | — | — | — | |
| OpenCLIPSupervision=(c) no supervision2023.06 | 61.3 | 34.4 | — | — | — | — | — | |
| DCTMBackbone=CAT-FCSS2019.08 | 61 | 38.1 | 72.1 | — | — | — | — |