Image Classification on ImageNet (test) (Corruption Accuracy and mCE)
77.5Clean AccuracyAugMix
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| AugMixBackbone=ResNet-50, Source Model Reference=[9]2021.12 | 77.5 | 48.3 | 65.3 | — | — | — | — | — | — | |
| PRIMEBackbone=ResNet-50, JSD Consistency Loss=false2021.12 | 77 | 55 | 57.5 | — | — | — | — | — | — | |
| DeepAugment (DA)Backbone=ResNet-50, Source Model Reference=[9]2021.12 | 76.7 | 52.6 | 60.4 | — | — | — | — | — | — | |
| StandardBackbone=ResNet-50, Source Model Reference=[9]2021.12 | 76.1 | 38.1 | 76.7 | — | — | — | — | — | — | |
| DA+AugMixBackbone=ResNet-502021.12 | 75.8 | 58.1 | 53.6 | — | — | — | — | — | — | |
| DA+PRIMEBackbone=ResNet-50, JSD Consistency Loss=false2021.12 | 75.5 | 59.9 | 51.3 | — | — | — | — | — | — | |
| Fast-AT†∞Training=Fast adversarial training (ℓ∞)2022.02 | — | — | — | 27.8 | 23.9 | 12.7 | 7 | 3.6 | 1 | |
| GAT-f†Backbone=ResNet50, Training=Proposed GAT approach2022.02 | — | — | — | 51 | 48.2 | 44.5 | 42.2 | 38.9 | 11.8 | |
| Madry†∞Backbone=ResNet50, Training=ℓ∞-robust training2022.02 | — | — | — | 38.1 | 33.9 | 21.9 | 14.4 | 9 | 2.8 | |
| Normal†Backbone=ResNet50, Training=Standard2022.02 | — | — | — | 50.9 | 45.8 | 33.4 | 25.8 | 21.1 | 0 |