Vision Foundation Models Bridge the Geometric Knowledge Across Domains for Long-Tailed Recognition

Yanbiao Ma, Licheng Jiao, Lingling Li, Bowei Liu, Fang Liu, Jiahao Wang, Shuyuan Yang, Xu Liu, Wenping Ma

Keywords: long-tailed classification, transfer of knowledg, distributed calibration, Vision foundation models

Abstract: Deep learning struggles to fully unleash its potential in scenarios with limited sample sizes, primarily because models fail to capture information beyond the observed domain when the number of samples from rare classes is limited. Therefore, restoring the true distribution of rare classes becomes a significant challenge. In this study, we discovered that vision foundation models can associate inter-class similarity with the similarity of geometric shapes of class distributions in cross-domain scenarios. Specifically, we observed that when two cross-domain classes are highly similar, their embedding distributions also exhibit similar geometric shapes and sizes. These phenomena only manifest when using foundation models to represent images. Our findings provide a foundation for leveraging geometric knowledge of existing data distributions to assist rare classes. Further, we propose the Geometrically Guided Uncertainty Representation (GUR) Layer tailored for long-tailed recognition tasks, aiming to calibrate and augment the embedding distribution of tail classes, thereby learning an unbiased MLP classifier. Across multiple long-tailed benchmark datasets, GUR significantly enhances the performance of vision foundation models and achieves state-of-the-art results on certain datasets. The success of GUR serves as a typical example of integrating and colliding foundation models with prior knowledge.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5986