Image Classification on ImageNet Natural Distribution Shifts (R, Sketch, A)
87.7Rendition (Acc)CLIP
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CLIPBackbone=ViT-L/14, Resolution=224x2242023.11 | 87.7 | 61.6 | 64.2 | 71.2 | |
| AugReg-ViT + DADBackbone=ViT-B/16, Resolution=224x2242023.11 | 65.1 | 46.1 | 31.8 | 47.7 | |
| AugReg-ViT + DAT + DADBackbone=ViT-B/16, Resolution=224x2242023.11 | 53.2 | 39.3 | 28.2 | 40.2 | |
| DADBackbone=ResNet50, Resolution=224x2242023.11 | 51.6 | 35.8 | 7.7 | 31.7 | |
| AugReg-ViT + Pyramid ATBackbone=ViT-B/16, Resolution=224x2242023.11 | 47.7 | 36.8 | 23 | 35.8 | |
| DAT + DADBackbone=ResNet50, Resolution=224x2242023.11 | 47.7 | 33.3 | 6.1 | 29 | |
| AugReg-ViT + DATBackbone=ViT-B/16, Resolution=224x2242023.11 | 47.3 | 34.8 | 30.2 | 37.4 | |
| AdvpropBackbone=ViT-B/16, Resolution=224x2242023.11 | 43.5 | 31.7 | 18.5 | 31.2 | |
| DATBackbone=ResNet50, Resolution=224x2242023.11 | 42 | 27.3 | 4.4 | 24.6 | |
| Fast AdvpropBackbone=ViT-B/16, Resolution=224x2242023.11 | 41.8 | 29.4 | 17.9 | 29.7 | |
| DebiasedBackbone=ResNet50, Resolution=224x2242023.11 | 40.8 | 28.4 | 3.5 | 24.2 | |
| DebiasedBackbone=ViT-B/16, Resolution=224x2242023.11 | 40.3 | 29.4 | 18.3 | 29.3 | |
| AugReg-ViTBackbone=ViT-B/16, Resolution=224x2242023.11 | 39.5 | 29.2 | 19 | 29.2 | |
| Pyramid ATBackbone=ResNet50, Resolution=224x2242023.11 | 38.9 | 23.8 | 3 | 21.9 | |
| AdvpropBackbone=ResNet50, Resolution=224x2242023.11 | 38.8 | 25.5 | 4.3 | 22.9 | |
| ResNet50Backbone=ResNet50, Resolution=224x2242023.11 | 36.1 | 24 | 0 | 20 | |
| ViTBackbone=ViT-B/16, Resolution=224x2242023.11 | 27.1 | 17.3 | 8 | 17.5 |