Reliable Boundary Samples-Based Proxy Pairs for Unsupervised Person Re-identification

Chang Zou, Zeqi Chen, Yuehu Liu, Chi Zhang

Abstract: Contrastive learning methods based on the memory bank have shown promising results for unsupervised person re-identification. However, most methods maintain a uni-proxy for each cluster in the memory that only describes the average information but can not represent the intra-class variation. As a result, contrastive learning based on the uni-proxy cannot effectively guide the model to reduce the variation. To address this issue, we maintain a proxy pair for each cluster updated by the least similar boundary sample pair since they concretely reveal the intra-class variation of the cluster. Through contrastive loss, the proxy updated based on one boundary sample generates strong pulls to another one and its surrounding samples due to the low similarity, and two proxies collaboratively form bidirectional strong pulls to effectively reduce intra-class variance. To mitigate the impact of boundary samples being noisy, we further propose local-global consistency guided label refinement, which utilizes local fine-grained cues to select reliable samples with high overlap in the local and global feature neighborhoods. Comprehensive experiments on Market-1501 and MSMT17 demonstrate that the proposed method surpasses state-of-the-art approaches.