Domain diversity based meta learning for continual person re-identification

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 3
• Main Authors: Liu, Zhaoshuo, Feng, Chaolu, Yu, Kun, Song, Jiangdian, Li, Wei
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2025; Springer Nature B.V
• Subjects: Computer Science; Learning; Original Article; Pattern Recognition; Continual learning; Domain diversity; Generalization; Person re-identification; Meta learning
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-025-01502-0

Continual person re-identification (CReID) aims to learn a unified model to cope with changing scenarios (e.g., from malls to streets and stations, etc). However, deploying the CReID model directly on unseen scenarios (which cannot be foreseen) outside the continual pipeline leads to a drop in generalization performance due to distribution shifts. In this paper, we design a meta learning paradigm with the seen and unseen domains to adapt the model to distribution shifts, where the unseen domains are produced by our carefully designed Domain Diversity (DD). Considering the knowledge encompassing both seen and unseen domains can fight against forgetting as well as improve generalization, we accumulate the learned and future knowledge corresponding to seen and unseen domains respectively through graph attention networks. Subsequently, we integrate the accumulated knowledge into meta learning steps to guide the training of the model, ensuring less forgetting and better generalization. Extensive experiments conducted on twelve datasets demonstrate the effectiveness of our method with superior performance in generalization and anti-forgetting. The code is available at https://github.com/DFLAG-NEU/MetaDD .