Structural consensus and collaborative correction for unsupervised visible-infrared person re-identification

Unsupervised visible-infrared person re-identification seeks to match pedestrian identities across modalities without manually annotated correspondences. While prevailing methods predominantly employ cluster-level associations for model learning, they encounter two fundamental limitations: the inherent sensitivity of clustering algorithms to feature noise leads to significant pseudo-label inaccuracies, and the over-reliance on cluster centroids overlooks fine-grained intra-cluster structures, thereby failing to resolve instance-level misalignments. In response to these challenges, we have developed a novel Neighborhood Correction Mechanism (NCM) that mitigates the aforementioned issues by incorporating a Neighborhood-Aware Pseudo-Label Refinement (NAPL) module. This module effectively denoises labels through neighborhood consistency, thereby enhancing label reliability. Furthermore, we introduce a Multi-scale Joint Matching (MJM) strategy that adaptively alternates between cluster-level matching and fine-grained reverse verification-based instance-level matching, significantly improving cross-modal semantic alignment. Extensive experiments validate the superiority of our approach, demonstrating substantial performance gains over existing state-of-the-art methods. Code is available at https://github.com/Lose-Code/NCM-net . • Neighborhood-aware pseudo-label correction for comprehensive cluster refinement in USL-VI-ReID. • Multi-level joint matching with reverse verification for robust cross-modal correspondence. • Extensive experiments show significant improvements in unsupervised visible-infrared person re-identification.