Reproduce “Diverse Representation Embedding for Lifelong Person Re-Identification” on OpenResearch

About this paper

Shiben Liu, Huijie Fan, Qiang Wang, Xiai Chen, Zhi Han, Yandong Tang

Researchers from the State Key Laboratory of Robotics, Chinese Academy of Sciences, developed DRE, a framework for Lifelong Person Re-Identification that leverages a pure transformer backbone with an Adaptive Constraint Module for diverse representation learning. This approach achieved average incremental gains of +6.6% mAP and +6.5% Rank-1 on seen datasets and demonstrated robust generalization on unseen and occluded datasets.

Key insights

• Transformer architectures, when adapted with features like Maximum Embedding and Multiple Class Tokens, can provide superior and more diverse representations for Lifelong Person Re-Identification than traditional CNNs.
• Ensuring richness and discriminability of representations through adaptive constraint mechanisms is crucial for effectively handling new data while preserving historical knowledge in continual learning.
• A dynamic teacher model, updated via Exponential Moving Average, combined with multi-level knowledge distillation and alignment strategies, effectively balances adaptation to new tasks and mitigation of catastrophic forgetting.

Summary by alphaXiv.