Cross-Modality Person Re-identification with Memory-Based Contrastive Embedding

Visible-infrared person re-identification (VI-ReID) aims to retrieve the images of same identity from RGB infrared image space, which is very important for real-world surveillance system. In practice, VI-ReID more challenging due heterogeneous modality discrepancy, further aggravates challenges traditional single-modality ReID problem, i.e., inter-class confusion and intra-class variations. this paper, we propose an aggregated memory-based cross-modality deep metric learning framework, benefits increasing number learned modality-aware modality-agnostic centroid proxies cluster contrast mutual information learning. Furthermore, suppress proposed alignment objective simultaneously utilizes both historical up-to-date enhanced association. Such training mechanism helps obtain hard positive references through increased diversity proxies, finally achieves stronger ``pulling close'' effect between features. Extensive experiment results demonstrate effectiveness method, surpassing state-of-the-art works significantly by a large margin on commonly used datasets.