Multi-granularity collaborative constraint feature alignment network for unsupervised person re-identification

Unsupervised Person Re-identification (UReID) aims to match pedestrians across non-overlapping cameras without annotations. Existing methods generally focus on mining multi-granularity feature complementarity, yet lack an explicit cross-granularity alignment mechanism, thus struggling to ensure stable semantic consistency across granularities during training. To address this issue, we propose a Multi-Granularity Collaborative Constraint Feature Alignment Network (MCCAN) for unsupervised person re-identification. As a multi-output, coarse-to-fine hierarchical alignment framework, MCCAN enables explicit alignment of multi-granularity features at all learning stages. Specifically, we design a multi-granularity constraint loss function to align local and global features explicitly, thus preserving their semantic consistency throughout training. Furthermore, we introduce a Center-Constrained Filtering (CCF) module, which leverages global feature centroids to constrain local features, achieving multi-granularity feature alignment within the memory bank. Extensive experiments on four public UReID datasets fully validate the effectiveness of the proposed method.