Dynamic adaptive multi-view contrastive learning for unsupervised person re-identification

Recent unsupervised person re-identification (Re-ID) methods leverage clustering to generate pseudo-labels for contrastive learning with a memory bank. However, camera variations introduce noise into these clustering-based pseudo-labels, and contrastive learning is hindered by inaccurate proxy construction, with hard pseudo-labels exhibiting inherent sensitivity to noise. This paper proposes a novel framework, Dynamic Adaptive Multi-view Contrastive Learning (DAMCL), to address these challenges. We introduce a Dynamic Adaptive Camera Jaccard (DACJ) distance to dynamically estimate and mitigate camera variations during each training epoch. Additionally, a Dynamic Adaptive Proxies (DAP) module, comprising Dynamic Optimal Cluster Proxies (DOCP) and Dynamic Instance Proxies (DIP), is proposed. Building on DACJ, DOCP forms the cluster proxy using the medoid of all cluster instances as its optimal feature representation. It aligns samples closely with their designated cluster proxy while distancing them from foreign proxies, using pseudo-labels generated by DBSCAN. Meanwhile, the DIP enhances clustering by leveraging global sample relationships. Finally, a Dynamic Adaptive Knowledge Distillation (DAKD) module is introduced to generate refined soft labels, improving robustness and accuracy. Comprehensive experiments confirm the efficiency of our approach.