A Memorizing and Generalizing Framework for Lifelong Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (ReID) models often struggle with catastrophic forgetting when trained incrementally across diverse domains.
• Existing methods lack effective strategies for continuous learning that balance generalization with knowledge retention.

Purpose of the Study:

• To propose a novel framework, MEmorizing and GEneralizing (MEGE), for lifelong person re-identification (LReID).
• To address the challenge of continuously training ReID models across multiple domains while ensuring generalization to unseen domains and preventing forgetting of seen domains.

Main Methods:

• The MEGE framework integrates Adaptive Knowledge Accumulation (AKA) using graph convolution networks for knowledge representation and operation.
• It also incorporates differentiable Ranking Consistency Distillation (RCD) to preserve ranking-specific knowledge.
• A new large-scale benchmark with practical evaluation protocols for LReID is introduced.

Main Results:

• The MEGE framework significantly improves performance on both seen and unseen domains under domain-incremental learning constraints.
• Experimental results demonstrate substantial performance gains compared to state-of-the-art methods.
• AKA effectively mitigates catastrophic forgetting and enhances generalization, while RCD further prevents knowledge loss.

Conclusions:

• The proposed MEGE framework offers an effective solution for the lifelong person re-identification problem.
• It successfully balances the need for generalization with the critical requirement of not forgetting previously acquired knowledge.
• The developed benchmark facilitates future research in LReID.