FedCAPR: Federated Camera-Aware Unsupervised Person Re-Identification with Identity-Distributed Equalization for Decentralized Data Clustering

Yu-Syuan Tseng, Tzu-Chin Hsu, Chih-Ting Liu, Shao-Yi Chien

Abstract: With the emergence of smart cities, person re-identification (re-ID) has become crucial for tracking individuals across surveillance footage. However, the acquisition of large-scale datasets with human annotations is not practical for real-world scenarios. Thus, purely unsupervised learning methods are widely employed in re-ID application. On the other hand, handling of sensitive personal data causes significant privacy concern. Fortunately, federated learning (FL) has been introduced to address privacy problems without exchanging data collected by each local client. To tackle both issues at the same time, previous works integrate unsupervised learning with FL on numerous re-ID benchmarks. Nevertheless, these existing methods face the difficulties due to the variance and complexity of data among clients, as well as the issue of information imbalance caused by traditional aggregation strategies such as FedAvg [??]. Therefore, we proposed FedCAPR, a novel federated unsupervised re-ID framework, which contains four principle features. First, we introduced a novel camera-aware contrastive loss to leverage camera information, significantly improving re-ID accuracy. Second, we incorporated a regularization penalty to ensure consistency and stability in the local federated optimization. Third, we developed the identity-distributed equalization (IDE) mechanism to address data heterogeneity across clients. Last, for FL aggregation, we introduced the cosine-distance weighted (CDW) method to guide the global model with a more effective aggregation during each round. Compared to state-of-the-art methods, FedCAPR achieves superior performance, up to a 23.4% improvement, on eight benchmark datasets. The source code will be made publicly available upon acceptance.

External IDs:dblp:conf/cvpr/TsengHLC25