Unsupervised Visible-Infrared Person Re-Identification via Fine-Grained Feature-Assisted Cross-Modality Alignment

Unsupervised visible–infrared person re-identification (UVI-ReID) aims to match person identities across visible and infrared camera views without relying on any manually annotated cross-modal labels. Existing UVI-ReID methods typically rely on global feature–based clustering and cross-modal alignment, but often neglect the role of fine-grained local features, leading to sub-optimal identity discrimination. To address these limitations, we propose a Fine-grained Feature-assisted Cross-modality Learning (FFCL) framework. Specifically, we introduce a Global–Local Joint Learning (GLJL) module that collaboratively learns identity-discriminative global and local representations via cluster-based contrastive learning during training. Meanwhile, we propose a Local-assisted Online Cross Pairing (LOCP) strategy, which leverages multi-part consensus to dynamically identify reliable cross-modal positive samples, alleviating the limitations of static clustering labels. Finally, during testing, we design a Local Consensus-based Cross-modal Retrieval (LCCR) method that fuses the consensus between local and global feature matchings to refine retrieval scores. Extensive experiments show that our approach outperforms existing unsupervised VI-ReID methods across multiple cross-modal datasets.