Optimizing Relevance Maps of Vision Transformers Improves Robustness

Open paper
Benchmark Model Rank Results
image-classification-on-objectnetAR-L (Opt Relevance)#33Top-1 Accuracy: 52.0Top-5 Accuracy: 73.5
image-classification-on-objectnetAR-B (Opt Relevance)#38Top-1 Accuracy: 47.1Top-5 Accuracy: 70
image-classification-on-objectnetAR-L#42Top-1 Accuracy: 46.5Top-5 Accuracy: 68.3
image-classification-on-objectnetViT-L (Opt Relevance)#43Top-1 Accuracy: 43.2Top-5 Accuracy: 65.8
image-classification-on-objectnetViT-B (Opt Relevance)#45Top-1 Accuracy: 42.2Top-5 Accuracy: 65.1
image-classification-on-objectnetAR-B#47Top-1 Accuracy: 41.4Top-5 Accuracy: 63.7
image-classification-on-objectnetAR-S (Opt Relevance)#50Top-1 Accuracy: 39.3Top-5 Accuracy: 61.7
image-classification-on-objectnetViT-L#53Top-1 Accuracy: 37.4Top-5 Accuracy: 59.5
image-classification-on-objectnetDeiT-L (Opt Relevance)#54Top-1 Accuracy: 36.3Top-5 Accuracy: 56.6
image-classification-on-objectnetViT-B#57Top-1 Accuracy: 35.1Top-5 Accuracy: 56.4
image-classification-on-objectnetAR-S#59Top-1 Accuracy: 34.3Top-5 Accuracy: 55.8
image-classification-on-objectnetDeiT-S (Opt Relevance)#62Top-1 Accuracy: 31.6Top-5 Accuracy: 53
image-classification-on-objectnetDeiT-L#64Top-1 Accuracy: 31.4Top-5 Accuracy: 48.5
image-classification-on-objectnetDeiT-S#77Top-1 Accuracy: 28.3Top-5 Accuracy: 47.3

Rank counts only results with a code link.