jianhongbai/bacon

Official implementation of "Towards Distribution-Agnostic Generalized Category Discovery" (NIPS 2023)

bacon's Issues

Question about Inference Result

Hi, thanks for your great work! I have a question about the inference result. In the paper, it is mentioned that the inference results are obtained by conduct k-means clustering on the post-backbone features. From my understanding, we can have multiple place to obtain the inference result: (1) clustering results from the post-backbone features, as used in the paper; (2) clustering results from the post-projector features in the contrastive-learning branch; (3) prediction results directly from the classifier in the pseudo-labeling branch, used in SimGCD paper.

Is there any justification why you choose the first (1) choice? Have tried the option (2) and (3)? Which one do you think will be better? We plan to include your paper as our baseline, you insights on these questions will be greatly appreciated 💌

Question about Fig. 1c

Hi, thanks for your interesting work. I have a question about Fig. 1c. Why labeled data for novel categories are encountered in real world while unlabeled samples are not available?