guojiajeremy/Dinomaly

about multi-class detection score

Closed this issue · 2 comments

xgzng commented

Hello, the result of running your code is the average of detecting each category separately. Will the scores of directly detecting all categories be different?

You raised a very keen question. The answer is NO. The average AUC of N classes does not equal to the AUC of all images of N classes mixed. However, all previous multi-class UAD methods implemented the former class-separate inference setting.

Actually, we have a another paper discussing it: https://arxiv.org/abs/2404.00724

xgzng commented

Okay, thank you for your quick reply.