Semantic Segmentation on ADE20K 150 classes (val)
40.7mIoURefineNet
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| RefineNetBackbone=ResNet-1522016.11 | 40.7 | — | — | — | |
| JAFARBackbone=DINOv2-ViT-S/14, Method Category=Task-Agnostic, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 40.49 | 74.92 | — | — | |
| RefineNetBackbone=ResNet-1012016.11 | 40.2 | — | — | — | |
| BilinearBackbone=DINOv2-ViT-S/14, Method Category=Training-free, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 39.23 | 73.69 | — | — | |
| DySampleBackbone=DINOv2-ViT-S/14, Method Category=Task-Dependent, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 38.99 | 73.62 | — | — | |
| ReSFuBackbone=DINOv2-ViT-S/14, Method Category=Task-Dependent, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 38.91 | 73.93 | — | — | |
| FeatUpBackbone=DINOv2-ViT-S/14, Method Category=Task-Agnostic, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 38.82 | 73.74 | — | — | |
| LIFTBackbone=DINOv2-ViT-S/14, Method Category=Task-Agnostic, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 38.73 | 73.69 | — | — | |
| CARAFEBackbone=DINOv2-ViT-S/14, Method Category=Task-Dependent, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 38.3 | 73.42 | — | — | |
| NearestBackbone=DINOv2-ViT-S/14, Method Category=Training-free, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 37.27 | 71.91 | — | — | |
| StridedBackbone=DINOv2-ViT-S/14, Method Category=Training-free, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 36.15 | 72.08 | — | — | |
| SAPABackbone=DINOv2-ViT-S/14, Method Category=Task-Dependent, Resolution=448x448, Evaluation Protocol=Linear Probing2025.06 | 35.87 | 71.85 | — | — | |
| Cascaded-DilatedNet2016.11 | 34.9 | — | — | — | |
| DilatedNet2016.11 | 32.3 | — | — | — | |
| FCN-8s2016.11 | 29.4 | — | — | — | |
| Cascaded-SegNet2016.11 | 27.5 | — | — | — | |
| Large Image (x8)Backbone=DINOv2-ViT-S/14, Method Category=Training-free, Resolution=Large Image (x8), Evaluation Protocol=Linear Probing2025.06 | 26.42 | 66.39 | — | — | |
| OpenSeg + Narr.zero-shot=true, segment_label=false, segment_mask=true2022.10 | 24.8 | — | — | — | |
| MaskCLIPCOCO Training Data=Masks + Labels2022.08 | 23.7 | — | — | — | |
| DeOPBackbone=R101c, Training dataset=COCO-Stuff-156, Decoupled network=true, Number of Passes=12023.04 | 22.9 | — | — | — | |
| MaskCLIP (MaskRCNN)COCO Training Data=Masks + Labels2022.08 | 22.4 | — | — | — | |
| SegNet2016.11 | 21.6 | — | — | — | |
| OpenSegzero-shot=true, segment_label=false, segment_mask=true2022.10 | 21.1 | — | — | — | |
| OpenSegCOCO Training Data=Masks + Captions2022.08 | 21.1 | — | — | — | |
| SimBaseBackbone=R101c, Training dataset=COCO-Stuff-156, Decoupled network=true, Number of Passes=N'2023.04 | 20.5 | — | — | — | |
| SimSegCOCO Training Data=Masks + Labels2022.08 | 20.5 | — | — | — | |
| LSeg+zero-shot=true, segment_label=true, segment_mask=true2022.10 | 18 | — | — | — | |
| LSeg+COCO Training Data=Masks + Labels2022.08 | 18 | — | — | — | |
| OpenSegBackbone=R101, Training dataset=COCO + Loc. Narr., Decoupled network=false, Number of Passes=12023.04 | 17.5 | — | — | — | |
| OpenSegBackbone=R101, Training dataset=COCO, Decoupled network=false, Number of Passes=12023.04 | 15.3 | — | — | — | |
| MaskCLIP w/o RMACOCO Training Data=Masks2022.08 | 14.9 | — | — | — | |
| CLIP BaselineCOCO Training Data=Masks2022.08 | 13.8 | — | — | — | |
| ALIGN w/ proposalzero-shot=true, segment_label=false, segment_mask=true2022.10 | 12.9 | — | — | — | |
| ALIGN w/ proposalsCOCO Training Data=Masks2022.08 | 12.9 | — | — | — | |
| ALIGNCOCO Training Data=None2022.08 | 10.7 | — | — | — | |
| BEiT-UperNetbackbone=BEiT-L†, crop size=6402021.12 | — | — | 57 | — | |
| FaPN-MaskFormerbackbone=Swin-L-FaPN†, crop size=6402021.12 | — | — | 55.2 | 56.7 | |
| Mask2Formerbackbone=R50, crop size=5122021.12 | — | — | 47.2 | 49.2 | |
| Mask2Formerbackbone=Swin-T, crop size=5122021.12 | — | — | 47.7 | 49.6 | |
| Mask2Formerbackbone=Swin-L†, crop size=6402021.12 | — | — | 56.1 | 57.3 | |
| Mask2Formerbackbone=Swin-L-FaPN†, crop size=6402021.12 | — | — | 56.4 | 57.7 | |
| MaskFormerbackbone=R50, crop size=5122021.12 | — | — | 44.5 | 46.7 | |
| MaskFormerbackbone=Swin-T, crop size=5122021.12 | — | — | 46.7 | 48.8 | |
| MaskFormerbackbone=Swin-L†, crop size=6402021.12 | — | — | 54.1 | 55.6 | |
| Swin-UperNetbackbone=Swin-T, crop size=5122021.12 | — | — | 46.1 | — |