Exploring Part Features for Unsupervised Visible-Infrared Person Re-Identification

Unsupervised visible-infrared person re-identification (USVI-ReID) is a challenging task that aims to retrieve images of the same person from different modalities without annotations. Existing works mainly focus on constructing cross-modality relationships with global features, the fine-grained part features remain unexplored, resulting in insufficient cross-modality learning. Therefore, we propose a Part-based Cross-Modality (PCM) learning framework to explore part features for USVI-ReID. Specifically, we first design a Part-integrated Dual-Contrastive (PDC) learning framework to obtain part features and learn discriminative information intramodality. Then, to associate samples from two modalities, we devise a Part-assisted Multiple Matching (PMM) module, which matches clusters with a weighted duplicated bipartite graph. Assisted by part features, the cost matrix for graph matching can be refined. Meanwhile, a Cross Alignment Learning (CAL) module is introduced to reduce modality discrepancy by aligning features at the granularity-level, memory-level and modality-level. Extensive experiments on two public datasets demonstrate the effectiveness of our proposed method.