Balancing Differential Discriminative Knowledge For Clothing-Irrelevant Lifelong Person Re-identification

Zhenyu Cui, Jiahuan Zhou, Yuxin Peng

Keywords: Person re-identification, Cloth-changing, Lifelong learning, Prototype learning

Abstract:

Lifelong person re-identification (L-ReID) focuses on learning sequentially collected datasets from different domains to match the same person. Advanced L-ReID methods typically balance the domain gap between different datasets via domain knowledge modeling, such as knowledge rectification or distribution prototyping. However, existing methods dismiss balancing discriminative knowledge within different datasets, resulting in conflicts when sequentially accumulating differential discriminative information in different datasets, e.g., sequentially learning cloth-changing/cloth-consistent knowledge simultaneously, which brings critical catastrophic forgetting problems of old discriminative knowledge. In this paper, we focus on a new but practical task called Cloth-Irrelevant Lifelong Per- sue, we proposed an Adaptive Discriminative Knowledge Consolidation (ADKC) framework to balance the discriminative information of different domains on L-ReID. Specifically, we propose a Selective Knowledge Forgetting (SKF) module to correct potential overfitting to specific discrimination (e.g., clothing information) based on new knowledge. In addition, we design a Selective Knowledge Retention (SKR) module to adaptively compensate for the potential lack of discriminative information based on old knowledge and accelerate differential discrimination into a unified framework. To validate our method, two CIL-ReID benchmarks are first established, while extensive experiments on the above two benchmark datasets demonstrate that our method leads to existing advanced methods in the CIL-ReID task.

Primary Area: transfer learning, meta learning, and lifelong 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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 653