Unsupervised visible-infrared person re-identification via locally reliable matching and global distribution...

Unsupervised Visible-Infrared Person Re-identification (US-VI-ReID) aims to match unlabeled pedestrian images across different modalities, and is a very challenging task. Realizing reliable cross-modality associations and mitigating cross-modality discrepancy are the key steps to obtain state-of-the-art recognition performance. However, existing methods usually neglect the reliability of different matching pairs, which introduces a large number of erroneous associations in the early training stage, thereby misleading model optimization and limiting the discriminative feature learning. Moreover, they fail to consider the structural consistency of feature distributions across different modalities, which is not conducive to the construction of reliable matching pairs and the mitigation of cross-modality discrepancies. To alleviate these issues, we propose a Locally Reliable Matching and Global Distribution Alignment (LRGD) method, which consists of three modules. First, we introduce a Reliability-aware Cross-modality Contrastive Learning (RCCL) module, which employs a fine-grained distance-guided matching matrix to dynamically assess the reliability of each cross-modality pair and assign adaptive weights. By emphasizing high-reliability pairs and suppressing low-quality ones, RCCL effectively mitigates the disturbance caused by noisy matching pairs and enhances the robustness of feature learning. Subsequently, we propose a Cross-modality Distribution Alignment(CMDA) method, which aligns the feature distributions of the visible and infrared modalities to enhance their structural consistency, thereby facilitating the generation of more reliable cross-modality matching pairs and mitigating the cross-modality discrepancy at the same time. Finally, considering that a single instance-level contrastive constraint is insufficient for learning reliable features, we design a Heterogeneous Integration and Homogeneous Cohesion (HIHC) module. By integrating cross-modality information, HIHC can generate an intermediate modality to mitigates the cross-modality discrepancy, which not only strengthens feature consistency within each class but also effectively reduces the distance between cross-modality positive pairs. Extensive experiments conducted on public datasets such as SYSU-MM01 and RegDB demonstrate that the proposed method achieves competitive cross-modality person reidentification performance, outperforming multiple state-of-the-art methods. The code is available at https://github.com/SuperGao12/LRGD.