Towards Better Generalization in Lifelong Person Re-Identification with Flatness-Aware Learning

Seungbin Hong, Sung Whan Yoon, Jae-Young SimOpenReview 2026 · Submission to ICLR 2026 · Rejected · OpenReview

Abstract

Lifelong person re-identification (LReID) requires models to continuously learn from sequentially arriving domains while retaining discriminative power for previously seen identities. A key challenge is to prevent catastrophic forgetting without access to old data, especially under exemplar-free constraints. In this paper, we propose a novel LReID method that unifies selective flatness-aware optimization, dual-model training, and model interpolation. Specifically, we maintain two separate models per task: a {stability model} trained with the distillation loss to retain the prior knowledge, and a {plasticity model} optimized solely for the current domain. To improve the performance of generalization and retention, we selectively apply Sharpness-Aware Minimization (SAM) only to the distillation loss, guiding the stability model toward flat and robust solutions. After task-specific training, these two models are fused through weight-space interpolation, producing a single model that balances stability and adaptability. The resulting model is used to initialize both branches for the next task, enabling continual knowledge integration. Our method is lightweight, modular, and readily compatible with existing LReID frameworks. Extensive experimental results consistently demonstrate that the proposed flat-minima-guided model fusion strategy consistently improves the overall performance of LReID.
BibTeX
@misc{hong2026towards,
  author = {Seungbin Hong and Sung Whan Yoon and Jae-Young Sim},
  title = {{Towards Better Generalization in Lifelong Person Re-Identification with Flatness-Aware Learning}},
  year = {2026},
  url = {https://openreview.net/forum?id=oNpSd5NS2H},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2026},
}
Figure