Metrics for CLIP-I and DreamSim

[chengyou-jia]

Thanks for your solid work!
I have some doubts about the calculations for CLIP-I and DreamSim. For instance, in the yaml under the first 'animals' entry, there are five data points. Are they calculated pairwise（C_5^2?） and then averaged? Also, did both metrics use CarveKit to clear the background, or was it just DreamSim? It would be great if the author could add these evaluations to the repository.

[byliutao]

Thank you for your interest in our work. We calculated the pairwise distances and then averaged. Both metrics use CarveKit to clear the background.