Object Segmentation on PartImageNet (val)
78.98mIoUCompositor
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Compositorbackbone=Swin-T, params=51M, training strategy=Models on parts and objects are trained jointly2023.06 | 78.98 | 87.8 | |
| MaskFormerbackbone=Swin-T, params=92M (46M×2), training strategy=Models on parts and objects are trained separately2023.06 | 77.92 | 87.44 | |
| MaskFormer-Dualbackbone=Swin-T, params=51M, training strategy=Models on parts and objects are trained jointly2023.06 | 77.24 | 87.12 | |
| SegFormerbackbone=MiT-B2, params=48M (24M×2), training strategy=Models on parts and objects are trained separately2023.06 | 74.55 | 85.24 | |
| Compositorbackbone=ResNet-50, params=50M, training strategy=Models on parts and objects are trained jointly2023.06 | 71.78 | 83.01 | |
| MaskFormer-Dualbackbone=ResNet-50, params=50M, training strategy=Models on parts and objects are trained jointly2023.06 | 70.44 | 81.81 | |
| MaskFormerbackbone=ResNet-50, params=90M (45M×2), training strategy=Models on parts and objects are trained separately2023.06 | 70.21 | 81.99 | |
| Deeplab v3+backbone=ResNet-50, params=42M, training strategy=Models on objects are trained with parts as deep supervision2023.06 | 69.82 | 81.96 | |
| Deeplab v3+backbone=ResNet-50, params=84M (42M×2), training strategy=Models on parts and objects are trained separately2023.06 | 68.38 | 81 |