FDGReID: Federated Domain Generalization for Person Re-Identification

Abstract

Person re-identification (Re-ID) has become a critical task in cross-camera retrieval systems. While deep learning-based approaches have made significant strides under controlled conditions, real-world deployment remains hindered by two major challenges: domain drift and data privacy. To address these challenges, we propose FDGReID, a novel federated learning framework designed to achieve domain generalization in person Re-ID without compromising user privacy. FDGReID introduces two core components: style information sharing (SIS) and viewpoint-aware contrastive learning (VCL). SIS diversifies stylistic exposure among distributed clients by sharing style representations during federated training, improving resilience to visual appearance changes. VCL, in contrast, mitigates spatial viewpoint shifts by enforcing identity consistency via contrastive objectives across varied perspectives at each client. Together, these modules enable FDGReID to learn robust, domain-invariant person representations without direct data exchange. We conduct extensive experiments on widely-used cross-domain Re-ID benchmarks, demonstrating that FDGReID consistently outperforms existing federated and generalizable Re-ID baselines. Moreover, it ensures strict data privacy compliance by keeping all raw images localized. Our results highlight FDGReID’s effectiveness and practicality in building scalable, privacy-preserving Re-ID systems for real-world applications.