baaivision / EVA

EVA Series: Visual Representation Fantasies from BAAI
MIT License
2.33k stars 167 forks source link

Robust accuracy on ObjectNet dataset. #136

Open sunrise6513 opened 10 months ago

sunrise6513 commented 10 months ago

It is shown that the robust accuracy of the fine-tuned EVA ViT on ObjectNet is obviously lower than that of the EVA CLIP model. Is that because the former is tested with imagenet 1k classes, while the latter being tested with ObjectNet classes?