Semantic Segmentation on Cityscapes (Mean IoU, Mean Pixel Acc)
72.29Mean IoUCORE-MTL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CORE-MTL2026.06 | 72.29 | 92.45 | |
| PaLoRAOptimizer=+MAdam2026.06 | 71.34 | 92.29 | |
| UWOptimizer=+MAdam2026.06 | 71.12 | 92.26 | |
| PaLoRAOptimizer=Base2026.06 | 71.11 | 92.21 | |
| LSOptimizer=+MAdam2026.06 | 70.89 | 92.14 | |
| DWAOptimizer=+MAdam2026.06 | 70.86 | 92.18 | |
| RepMTL2026.06 | 70.79 | 91.84 | |
| IMTLOptimizer=Base2026.06 | 70.77 | 92.12 | |
| CAGradOptimizer=+MAdam2026.06 | 70.71 | 91.63 | |
| PaMaLOptimizer=+MAdam2026.06 | 70.59 | 92.17 | |
| KAConvNet-LBackbone=KAConvNet-L, Segmentation Head=PSPNet [49]2026.04 | 70.58 | 79.51 | |
| IMTLOptimizer=+MAdam2026.06 | 70.49 | 92.02 | |
| PaMaLOptimizer=Base2026.06 | 70.35 | 91.99 | |
| HVT BaseBackbone=HVT Base, Segmentation Head=PSPNet [49]2026.04 | 70.32 | 79.25 | |
| MTAN2026.06 | 70.23 | 92.06 | |
| MobileOne-S4Backbone=MobileOne-S4, Segmentation Head=PSPNet [49]2026.04 | 70.21 | 79.36 | |
| UWOptimizer=Base2026.06 | 70.2 | 91.93 | |
| GradNorm2026.06 | 70.15 | 91.97 | |
| LSOptimizer=Base2026.06 | 70.12 | 91.9 | |
| DWAOptimizer=Base2026.06 | 70.1 | 91.89 | |
| PCGrad2026.06 | 69.98 | 91.87 | |
| RLW2026.06 | 69.95 | 91.85 | |
| RepVGG-A2Backbone=RepVGG-A2, Segmentation Head=PSPNet [49]2026.04 | 69.94 | 78.89 | |
| FairGrad2026.06 | 69.86 | 91.94 | |
| ExcessMTL2026.06 | 69.69 | 91.95 | |
| Equal Weighting2026.06 | 69.62 | 91.92 | |
| STCH2026.06 | 69.52 | 91.8 | |
| CAGradOptimizer=Base2026.06 | 69.23 | 91.61 | |
| KAConvNet-BBackbone=KAConvNet-B, Segmentation Head=PSPNet [49]2026.04 | 69.2 | 78.7 | |
| ViL-Tiny-APEBackbone=ViL-Tiny-APE, Segmentation Head=PSPNet [49]2026.04 | 69.1 | 78.52 | |
| LocalViT-TBackbone=LocalViT-T, Segmentation Head=PSPNet [49]2026.04 | 69.07 | 78.55 | |
| MOML2026.06 | 68.76 | 91.46 | |
| Single Task2026.06 | 68.69 | 91.44 | |
| GhostNet1.3Backbone=GhostNet1.3, Segmentation Head=PSPNet [49]2026.04 | 68.66 | 78.47 | |
| RepVGG-A1Backbone=RepVGG-A1, Segmentation Head=PSPNet [49]2026.04 | 68.52 | 78.33 | |
| ResNet-18Backbone=ResNet-18, Segmentation Head=PSPNet [49]2026.04 | 68.02 | 77.9 | |
| KAConvNet-SBackbone=KAConvNet-S, Segmentation Head=PSPNet [49]2026.04 | 65.32 | 76.06 | |
| MobileV3-L 0.75Backbone=MobileV3-L 0.75, Segmentation Head=PSPNet [49]2026.04 | 65.19 | 76.01 | |
| DeiT-TinyBackbone=DeiT-Tiny, Segmentation Head=PSPNet [49]2026.04 | 64.95 | 75.9 | |
| Ours (WRC)Evaluation Protocol=Linear Probing, Upsampling=2x2026.05 | 62.02 | 93.51 | |
| LoftUpEvaluation Protocol=Linear Probing, Upsampling=Default2026.05 | 61.94 | 93.51 | |
| JAFAREvaluation Protocol=Linear Probing, Upsampling=Default2026.05 | 61.87 | 93.51 | |
| AnyUpEvaluation Protocol=Linear Probing, Upsampling=Default2026.05 | 61.45 | 93.44 | |
| FeatUpEvaluation Protocol=Linear Probing, Upsampling=Default2026.05 | 60.41 | 93.16 | |
| LiFTEvaluation Protocol=Linear Probing, Upsampling=Default2026.05 | 60.08 | 92.9 | |
| BilinearEvaluation Protocol=Linear Probing, Upsampling=Default2026.05 | 59.74 | 92.55 | |
| DINOArch=ViT-B, Pretrain=SSv2, Evaluation Framework=UperNet2026.06 | 39.97 | 43.09 | |
| DINOArch=ViT-S, Pretrain=SSv2, Evaluation Framework=UperNet2026.06 | 39.85 | 45.68 | |
| iBOTArch=ViT-S, Pretrain=SSv2, Evaluation Framework=UperNet2026.06 | 39.34 | 45.36 | |
| iBOTArch=ViT-B, Pretrain=SSv2, Evaluation Framework=UperNet2026.06 | 38.94 | 44.31 | |
| TDVArch=ViT-S, Pretrain=SSv2, Evaluation Framework=UperNet2026.06 | 37.54 | 43.09 | |
| TDVArch=ViT-B, Pretrain=SSv2, Evaluation Framework=UperNet2026.06 | 36.21 | 42.59 | |
| CEPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.25/2552026.05 | 34.9 | 72 | |
| SegPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.25/2552026.05 | 34.1 | 66.9 | |
| MaskedPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.25/2552026.05 | 32.8 | 63.7 | |
| TsallisPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.25/2552026.05 | 32.8 | 62.9 | |
| JSPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.25/2552026.05 | 32.4 | 66 | |
| CosPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.25/2552026.05 | 32.1 | 64.7 | |
| CEPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.5/2552026.05 | 20 | 42.2 | |
| SegPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.5/2552026.05 | 18.3 | 36.2 | |
| JSPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.5/2552026.05 | 14.3 | 29.2 | |
| TsallisPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.5/2552026.05 | 14.1 | 27.3 | |
| CosPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.5/2552026.05 | 14 | 27.1 | |
| MaskedPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=0.5/2552026.05 | 13.9 | 26.2 | |
| CEPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.25/2552026.05 | 12.9 | 22.1 | |
| SegPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.25/2552026.05 | 9.2 | 14.1 | |
| JSPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.25/2552026.05 | 6.5 | 9 | |
| CEPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=1/2552026.05 | 6.4 | 16.2 | |
| CosPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.25/2552026.05 | 5.8 | 7.9 | |
| SegPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=1/2552026.05 | 5.8 | 12.4 | |
| TsallisPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.25/2552026.05 | 5.6 | 7.2 | |
| MaskedPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.25/2552026.05 | 5 | 6.5 | |
| TsallisPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=1/2552026.05 | 2.1 | 4.1 | |
| JSPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=1/2552026.05 | 1.8 | 3.8 | |
| MaskedPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=1/2552026.05 | 1.6 | 2.7 | |
| CEPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.5/2552026.05 | 1.5 | 5.3 | |
| CosPGDbackbone=PSPNet+ResNet-50, training_strategy=DDCAT, epsilon=1/2552026.05 | 1.5 | 3 | |
| CEPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=1/2552026.05 | 1 | 3.9 | |
| SegPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.5/2552026.05 | 0.3 | 1 | |
| SegPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=1/2552026.05 | 0.3 | 0.5 | |
| CosPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.5/2552026.05 | 0.2 | 0.5 | |
| JSPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.5/2552026.05 | 0.2 | 0.7 | |
| MaskedPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.5/2552026.05 | 0.1 | 0.2 | |
| TsallisPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=0.5/2552026.05 | 0.1 | 0.2 | |
| CosPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=1/2552026.05 | 0.1 | 0.2 | |
| JSPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=1/2552026.05 | 0.1 | 0.1 | |
| MaskedPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=1/2552026.05 | 0 | 0 | |
| TsallisPGDbackbone=PSPNet+ResNet-50, training_strategy=clean training, epsilon=1/2552026.05 | 0 | 0 |