Perturbation Detection for Semantic Segmentation on ADE20K (val)
37.7mIoU (Clean)Entropy
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| EntropyBackbone=SegFormer MiT-B0, Training Dataset=D_ADE20K_train, Perturbation strengths (epsilon)=1, 2, 4, 8, 162026.06 | 37.7 | 73.9 | 71.2 | 47.3 | 50 | 52.3 | 58.9 | |
| Max posteriorBackbone=SegFormer MiT-B0, Training Dataset=D_ADE20K_train, Perturbation strengths (epsilon)=1, 2, 4, 8, 162026.06 | 37.7 | 72.7 | 70 | 46.9 | 50 | 52 | 58.3 | |
| Error (17)Backbone=SegFormer MiT-B0, Training Dataset=D_ADE20K_train, Perturbation strengths (epsilon)=1, 2, 4, 8, 162026.06 | 35.2 | 61.9 | 64.2 | 51.1 | 50 | 52.9 | 56 | |
| Reg. (18)Backbone=SegFormer MiT-B0, Training Dataset=D_ADE20K_train, Perturbation strengths (epsilon)=1, 2, 4, 8, 162026.06 | 35.2 | 47 | 50.7 | 52.9 | 50 | 49.7 | 50.1 | |
| Quantile (19)Backbone=SegFormer MiT-B0, Training Dataset=D_ADE20K_train, Perturbation strengths (epsilon)=1, 2, 4, 8, 162026.06 | 35.2 | 53.9 | 60.4 | 56.3 | 50 | 52.5 | 54.6 |