Semantic Segmentation on PASCAL VOC
0.8952mIoUDINOv2
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DINOv2Arch.=ViT-B*, SSL Type Target=GL+DE, SSL Type Method=DIST+MIM, Evaluation Protocol=linear probing2026.02 | 0.8952 | — | |
| DINOv2Arch.=ViT-L*, SSL Type Target=GL+DE, SSL Type Method=DIST+MIM, Evaluation Protocol=linear probing2026.02 | 0.8919 | — | |
| MaskCLIP+paradigm=transductive2021.12 | 0.888 | — | |
| Fully Sup.2021.12 | 0.886 | — | |
| DINOv3-B/16Protocol=Linear probing, Backbone=B/16, Learning Paradigm=Self-supervised2026.05 | 0.875 | — | |
| VECA-B/16Protocol=Linear probing, Backbone=B/16, Core budget (C)=642026.05 | 0.8707 | — | |
| STELLARArch.=ViT-L, SSL Type Target=SPARSE, SSL Type Method=INV+REC, Evaluation Protocol=linear probing2026.02 | 0.859 | — | |
| AM-RADIOv2.5-B/16Protocol=Linear probing, Backbone=B/16, Learning Paradigm=Agglomerative2026.05 | 0.8573 | — | |
| Omnivorous ViT-B/14readout=DPT2026.02 | 0.857 | — | |
| STELLARArch.=ViT-H, SSL Type Target=SPARSE, SSL Type Method=INV+REC, Evaluation Protocol=linear probing2026.02 | 0.8566 | — | |
| DINOv2 ViT-B/14readout=DPT2026.02 | 0.855 | — | |
| DINOv2-B/14Protocol=Linear probing, Backbone=B/14, Learning Paradigm=Self-supervised2026.05 | 0.8446 | — | |
| JAFARLinear probing=1x1 convolution layer, Input resolution=448x448, Upsampling factor=14x or 16x2025.10 | 0.8436 | 96.22 | |
| DINOv2-reg-B/14Protocol=Linear probing, Backbone=B/14, Learning Paradigm=Self-supervised2026.05 | 0.8423 | — | |
| AnyUpLinear probing=1x1 convolution layer, Input resolution=448x448, Upsampling factor=14x or 16x2025.10 | 0.84 | 96.19 | |
| VECA-B/16Protocol=Linear probing, Backbone=B/16, Core budget (C)=82026.05 | 0.8384 | — | |
| LoftUpLinear probing=1x1 convolution layer, Input resolution=448x448, Upsampling factor=14x or 16x2025.10 | 0.8369 | 96.11 | |
| FeatUpLinear probing=1x1 convolution layer, Input resolution=448x448, Upsampling factor=14x or 16x2025.10 | 0.8337 | 96.04 | |
| MuRF (Ours)Arch.=ViT-B/142026.03 | 0.831 | — | |
| PSPNet + L2T-DLNloss_type=DLN2023.10 | 0.829 | — | |
| STRICTparadigm=transductive2021.12 | 0.827 | — | |
| PSPNetloss_type=original2023.10 | 0.826 | — | |
| Omnivorous ViT-B/14readout=Linear2026.02 | 0.826 | — | |
| iBOTArch.=ViT-L/162026.03 | 0.823 | — | |
| STELLARArch.=ViT-B, SSL Type Target=SPARSE, SSL Type Method=INV+REC, Evaluation Protocol=linear probing2026.02 | 0.8183 | — | |
| High ResolutionArch.=ViT-B/142026.03 | 0.818 | — | |
| BilinearLinear probing=1x1 convolution layer, Input resolution=448x448, Upsampling factor=14x or 16x2025.10 | 0.8143 | 95.38 | |
| DINOv2 ViT-B/14readout=Linear2026.02 | 0.814 | — | |
| SiameseIMArch.=ViT-B, SSL Type Target=DENSE, SSL Type Method=LAT-MIM, Evaluation Protocol=linear probing2026.02 | 0.8138 | — | |
| DINOArch.=ViT-B, SSL Type Target=GLOBAL, SSL Type Method=DISTILL, Evaluation Protocol=linear probing2026.02 | 0.7929 | — | |
| CaGNetparadigm=transductive2021.12 | 0.786 | — | |
| SimCLRArchitecture=DeepLab-v32021.10 | 0.785 | — | |
| Medium ResolutionArch.=ViT-B/142026.03 | 0.785 | — | |
| CaGNetparadigm=inductive2021.12 | 0.784 | — | |
| DINO-BEncoder=DINO-B, Model=Teacher (frozen DINO), Repr.=teacher, Evaluation Protocol=Linear probe2026.05 | 0.784 | — | |
| DINO-BRepresentation=teacher, Evaluation protocol=Linear probing2026.05 | 0.784 | — | |
| DCNv2Backbone=ResNet-101, offset & modulation pretraining=ImageNet2018.11 | 0.783 | — | |
| MAEArch.=ViT-H, SSL Type Target=DENSE, SSL Type Method=PIX MIM, Evaluation Protocol=linear probing2026.02 | 0.7807 | — | |
| SPNet-Cparadigm=inductive2021.12 | 0.78 | — | |
| ZS3Netparadigm=transductive2021.12 | 0.78 | — | |
| SPNetparadigm=transductive2021.12 | 0.778 | — | |
| MAEArch.=ViT-L, SSL Type Target=DENSE, SSL Type Method=PIX MIM, Evaluation Protocol=linear probing2026.02 | 0.7779 | — | |
| iBOTArch.=ViT-L, SSL Type Target=GL+DE, SSL Type Method=DIST+MIM, Evaluation Protocol=linear probing2026.02 | 0.7757 | — | |
| TWISTArchitecture=DeepLab-v32021.10 | 0.773 | — | |
| ZS3Netparadigm=inductive2021.12 | 0.773 | — | |
| SwAVArchitecture=DeepLab-v32021.10 | 0.772 | — | |
| iBOTArch.=ViT-B, SSL Type Target=GL+DE, SSL Type Method=DIST+MIM, Evaluation Protocol=linear probing2026.02 | 0.7706 | — | |
| SupArchitecture=DeepLab-v32021.10 | 0.766 | — | |
| MAEArch.=ViT-B, SSL Type Target=DENSE, SSL Type Method=PIX MIM, Evaluation Protocol=linear probing2026.02 | 0.7643 | — | |
| DINOArchitecture=DeepLab-v32021.10 | 0.764 | — | |
| DINO-LEncoder=DINO-L, Model=Teacher (frozen DINO), Repr.=teacher, Evaluation Protocol=Linear probe2026.05 | 0.762 | — | |
| DCNv2Backbone=ResNet-101, offset & modulation pretraining=none2018.11 | 0.761 | — | |
| DC-v2Architecture=DeepLab-v32021.10 | 0.76 | — | |
| NAT-M1#MAdds=225M2020.05 | 0.759 | — | |
| SPNetparadigm=inductive2021.12 | 0.758 | — | |
| SigLIP 2-B/16Protocol=Linear probing, Backbone=B/16, Learning Paradigm=Weakly Supervised2026.05 | 0.7541 | — | |
| SlotConEpochs=200, w/ FPN=false, Obj. Prior=false2022.05 | 0.75 | — | |
| SlotConPre-training Dataset=ImageNet, Epochs=2002022.05 | 0.75 | — | |
| ImageNet supervisedpre-train=super. IN-1M2019.11 | 0.744 | — | |
| supervisedEpochs=100, w/ FPN=false, Obj. Prior=false2022.05 | 0.744 | — | |
| I-JEPAArch.=ViT-H, SSL Type Target=DENSE, SSL Type Method=LAT-MIM, Evaluation Protocol=linear probing2026.02 | 0.7413 | — | |
| MoCo v3Arch.=ViT-B, SSL Type Target=GLOBAL, SSL Type Method=CONTR., Evaluation Protocol=linear probing2026.02 | 0.7408 | — | |
| PixPro+Epochs=100, w/ FPN=false, Obj. Prior=false2022.05 | 0.739 | — | |
| SlotConPre-training Dataset=COCO+, Epochs=8002022.05 | 0.739 | — | |
| MobileNetV3#MAdds=219M2020.05 | 0.738 | — | |
| MoCo v2+Epochs=800, w/ FPN=false, Obj. Prior=false2022.05 | 0.737 | — | |
| MoCopre-train=MoCo IG-1B2019.11 | 0.736 | — | |
| FBNetV2#MAdds=238M2020.05 | 0.736 | — | |
| MoCo + CC + A+ + kNNPre-training Dataset=COCO, Pre-training Epochs=800, Constrained multi-crop=true, Stronger augmentations=true, kNN nearest neighbors=true2021.06 | 0.735 | — | |
| TWISTArchitecture=FCN-FPN2021.10 | 0.733 | — | |
| TWISTBackbone=FCN-FPN2021.10 | 0.733 | — | |
| DenseCLPre-training Dataset=COCO, Pre-training Epochs=8002021.06 | 0.732 | — | |
| SlotConEpochs=100, w/ FPN=false, Obj. Prior=false2022.05 | 0.731 | — | |
| SlotConPre-training Dataset=ImageNet, Epochs=1002022.05 | 0.731 | — | |
| Moco-v2Architecture=DeepLab-v32021.10 | 0.729 | — | |
| InsLoc+Epochs=200, w/ FPN=true, Obj. Prior=false2022.05 | 0.729 | — | |
| CLIP-B/16Protocol=Linear probing, Backbone=B/16, Learning Paradigm=Weakly Supervised2026.05 | 0.7287 | — | |
| SimCLRArchitecture=FCN-FPN2021.10 | 0.728 | — | |
| SimCLRBackbone=FCN-FPN2021.10 | 0.728 | — | |
| DenseCL+Epochs=200, w/ FPN=false, Obj. Prior=false2022.05 | 0.728 | — | |
| VirTexPre-training Dataset=COCO captions2021.06 | 0.727 | — | |
| MoCo + CC + A+Pre-training Dataset=COCO, Pre-training Epochs=800, Constrained multi-crop=true, Stronger augmentations=true2021.06 | 0.727 | — | |
| DetCo+Epochs=200, w/ FPN=false, Obj. Prior=false2022.05 | 0.726 | — | |
| DetConEpochs=200, w/ FPN=false, Obj. Prior=true2022.05 | 0.726 | — | |
| MoCopre-train=MoCo IN-1M2019.11 | 0.725 | — | |
| MoCo + CCPre-training Dataset=COCO, Pre-training Epochs=800, Constrained multi-crop=true2021.06 | 0.722 | — | |
| DC-v2Architecture=FCN-FPN2021.10 | 0.721 | — | |
| DC-v2Backbone=FCN-FPN2021.10 | 0.721 | — | |
| regularBackbone=ResNet-101, offset & modulation pretraining=none2018.11 | 0.72 | — | |
| SwAVArchitecture=FCN-FPN2021.10 | 0.719 | — | |
| DINOArchitecture=FCN-FPN2021.10 | 0.719 | — | |
| SwAVBackbone=FCN-FPN2021.10 | 0.719 | — | |
| DINOBackbone=FCN-FPN2021.10 | 0.719 | — | |
| SoCo+Epochs=100, w/ FPN=true, Obj. Prior=true2022.05 | 0.719 | — | |
| SlotConPre-training Dataset=COCO, Epochs=8002022.05 | 0.716 | — | |
| DFNCLIP-B/16Protocol=Linear probing, Backbone=B/16, Learning Paradigm=Weakly Supervised2026.05 | 0.7158 | — | |
| OpenCLIPArch.=ViT-G/142026.03 | 0.714 | — | |
| MoCoPre-training Dataset=COCO, Pre-training Epochs=8002021.06 | 0.711 | — | |
| DenseCLArch.=RN-50, SSL Type Target=DENSE, SSL Type Method=CONTR., Evaluation Protocol=linear probing2026.02 | 0.7095 | — | |
| SetSimArchitecture=FCN (R50), Evaluation Protocol=Fine-tuning, Pre-training Epochs=2002021.07 | 0.709 | — |