Generalized Category Discovery under the Long-Tailed Distribution
Abstract: Generalized Category Discovery (GCD) is a challenging problem, which aims at discovering novel categories in unlabelled data by transferring the knowledge from the labelled data. Existing methods often assume a uniform distribution of categories, which is not representative of real-world data that typically exhibits a long-tailed distribution. In this paper, we address the problem of GCD under a long-tailed distribution. Our approach introduces a novel framework that tackles the challenges of biased classifier learning and imprecise class number estimation. We propose adaptive sample selection based on confidence and density, balancing the model's training distribution and mitigating bias. Additionally, we present a density-peak-based method for accurate class number estimation in long-tailed settings. Experimental results demonstrate the effectiveness of our approach in discovering novel categories and outperforming state-of-the-art methods.
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=fLXiCTre3j
Changes Since Last Submission: We have changed the font to the correct one for TMLR.
Assigned Action Editor: ~Gang_Niu1
Submission Number: 2740
Loading