Semantic Segmentation on ADE20K (mIoU SS and MS)
59.09mIoU (SS)SENSE
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SENSESynthetic data volume=2×2026.05 | 59.09 | 60.81 | |
| SegGenSynthetic data volume=50×2026.05 | 57.4 | 58.7 | |
| FreeMaskSynthetic data volume=20×2026.05 | 56.4 | — | |
| VMamba-BBackbone=VMamba-B, Param (M)=122, FLOP (G)=1170, Framework=UperNet, Input Resolution=512×20482025.03 | 51 | 51.6 | |
| VMamba-SBackbone=VMamba-S, Param (M)=82, FLOP (G)=1028, Framework=UperNet, Input Resolution=512×20482025.03 | 50.6 | 51.2 | |
| ConvNeXt-SBackbone=ConvNeXt-S, Param (M)=82, FLOP (G)=1027, Framework=UperNet, Input Resolution=512×20482025.03 | 48.7 | 49.6 | |
| Full fine-tuningBackbone=ViT-L, Params.(M)=318.312026.06 | 48.31 | 50.07 | |
| Swin-BBackbone=Swin-B, Param (M)=121, FLOP (G)=1188, Framework=UperNet, Input Resolution=512×20482025.03 | 48.1 | — | |
| VMamba-TBackbone=VMamba-T, Param (M)=62, FLOP (G)=949, Framework=UperNet, Input Resolution=512×20482025.03 | 47.9 | 48.8 | |
| Swin-SBackbone=Swin-S, Param (M)=81, FLOP (G)=1039, Framework=UperNet, Input Resolution=512×20482025.03 | 47.6 | 49.5 | |
| vGamba-LBackbone=vGamba-L, Param (M)=71, FLOP (G)=1019, Framework=UperNet, Input Resolution=512×20482025.03 | 46.9 | 47.8 | |
| vGamba-BBackbone=vGamba-B, Param (M)=55, FLOP (G)=941, Framework=UperNet, Input Resolution=512×20482025.03 | 46.6 | 46.7 | |
| Eff.VMamba-BBackbone=Eff.VMamba-B, Param (M)=65, FLOP (G)=930, Framework=UperNet, Input Resolution=512×20482025.03 | 46.5 | 47.3 | |
| ConvNeXt-TBackbone=ConvNeXt-T, Param (M)=60, FLOP (G)=939, Framework=UperNet, Input Resolution=512×20482025.03 | 46 | 46.7 | |
| DeiT-B + MLNBackbone=DeiT-B + MLN, Param (M)=144, FLOP (G)=2007, Framework=UperNet, Input Resolution=512×20482025.03 | 45.5 | 47.2 | |
| HyperAdapterBackbone=ViT-L, Params.(M)=14.722026.06 | 45.2 | 46.95 | |
| Vim-SBackbone=Vim-S, Param (M)=46, Framework=UperNet, Input Resolution=512×20482025.03 | 44.9 | — | |
| Swin-TBackbone=Swin-T, Param (M)=60, FLOP (G)=945, Framework=UperNet, Input Resolution=512×20482025.03 | 44.5 | 45.8 | |
| RepAdapterBackbone=ViT-L, Params.(M)=13.822026.06 | 44.44 | 46.17 | |
| DeiT-S + MLNBackbone=DeiT-S + MLN, Param (M)=58, FLOP (G)=1217, Framework=UperNet, Input Resolution=512×20482025.03 | 43.8 | 45.1 | |
| ResNet-101Backbone=ResNet-101, Param (M)=85, FLOP (G)=1030, Framework=UperNet, Input Resolution=512×20482025.03 | 42.9 | 44 | |
| VPTBackbone=ViT-L, Params.(M)=13.432026.06 | 42.11 | 44.06 | |
| ResNet-50Backbone=ResNet-50, Param (M)=67, FLOP (G)=953, Framework=UperNet, Input Resolution=512×20482025.03 | 42.1 | 42.8 | |
| Eff.VMamba-SBackbone=Eff.VMamba-S, Param (M)=29, FLOP (G)=505, Framework=UperNet, Input Resolution=512×20482025.03 | 41.5 | 42.1 | |
| Eff.VMamba-TBackbone=Eff.VMamba-T, Param (M)=14, FLOP (G)=230, Framework=UperNet, Input Resolution=512×20482025.03 | 38.9 | 39.3 | |
| Linear probingBackbone=ViT-L, Params.(M)=13.182026.06 | 35.12 | 37.46 |