Image Classification on ImageNet Robustness Variants (Adversarial, Rendition, Sketch) V2 (test)
84.21Accuracy (ID)Adam-SPD
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Adam-SPDBackbone=CLIP ViT-Base, Protocol=Fine-Tuning, Regularization=Selective Projection Decay2024.11 | 84.21 | 74.83 | 25.42 | 49.09 | 51.18 | 50.13 | 56.95 | |
| FTPBackbone=CLIP ViT-Base, Protocol=Fine-Tuning, Regularization=FTP2024.11 | 84.19 | 74.64 | 26.5 | 47.23 | 50.23 | 49.65 | 56.56 | |
| SAMBackbone=CLIP ViT-Base, Protocol=Fine-Tuning, Optimizer=SAM2024.11 | 83.67 | 73.66 | 20.48 | 42.98 | 45.7 | 45.71 | 53.3 | |
| Vanilla FTBackbone=CLIP ViT-Base, Protocol=Fine-Tuning2024.11 | 83.66 | 73.82 | 21.4 | 43.06 | 45.52 | 46.98 | 54.29 | |
| L2-SPBackbone=CLIP ViT-Base, Protocol=Fine-Tuning, Regularization=L2-SP2024.11 | 83.44 | 73.2 | 20.55 | 43.89 | 46.6 | 46.06 | 53.54 | |
| LP-FTBackbone=CLIP ViT-Base, Protocol=Linear Probing then Fine-Tuning2024.11 | 82.99 | 72.96 | 21.08 | 44.65 | 47.56 | 46.56 | 53.85 | |
| WiSE-SPDBackbone=CLIP ViT-Base, Protocol=Weight Interpolation, Regularization=Selective Projection Decay2024.11 | 81.7 | 73.29 | 34.37 | 63.69 | 54.55 | 56.48 | 61.52 | |
| WiSE-FTBackbone=CLIP ViT-Base, Protocol=Weight Interpolation2024.11 | 80.94 | 72.47 | 33.18 | 63.33 | 54.2 | 55.58 | 60.82 | |
| Linear Prob.Backbone=CLIP ViT-Base, Protocol=Linear Probing2024.11 | 78.25 | 67.68 | 26.54 | 52.57 | 48.26 | 48.76 | 54.66 | |
| SpecPLbaseline=CoOp2026.05 | 72.12 | 65.25 | 50.45 | 76.3 | 48.25 | — | 60.06 | |
| SpecPLbaseline=MMRL2026.05 | 72.07 | 64.6 | 51.43 | 77.67 | 49.33 | — | 60.76 | |
| MMRL2026.05 | 72.03 | 64.47 | 51.2 | 77.53 | 49.17 | — | 60.59 | |
| CoOp2026.05 | 71.51 | 64.2 | 49.71 | 75.21 | 47.99 | — | 59.28 | |
| Zero-ShotBackbone=CLIP ViT-Base, Protocol=Zero-Shot2024.11 | 67.68 | 61.41 | 30.6 | 56.77 | 45.53 | 48.58 | 52.4 |