Beyond a strong baseline: cross-modality contrastive learning for visible-infrared person re-identification

Cross-modality pedestrian image matching, which entails the matching of visible and infrared images, is a vital area in person re-identification (reID) due to its potential to facilitate person retrieval across a spectrum of lighting conditions. Despite its importance, this task presents considerable challenges stemming from two significant areas: cross-modality discrepancies due to the different imaging principles of spectrum cameras and within-class variations caused by the diverse viewpoints of large-scale distributed surveillance cameras. Unfortunately, the existing literature provides limited insights into effectively mitigating these issues, signifying a crucial research gap. In response to this, the present paper makes two primary contributions. First, we conduct a comprehensive study of training methodologies and subsequently present a strong baseline network designed specifically to address the complexities of the visible-infrared person reID task. This strong baseline network is paramount to the advancement of the field and to ensure the fair evaluation of algorithmic effectiveness. Second, we propose the Cross-Modality Contrastive Learning (CMCL) scheme, a novel approach to address the cross-modality discrepancies and enhance the quality of image embeddings across both modalities. CMCL incorporates intra-modality and inter-modality contrastive loss components, designed to improve the matching quality across the modalities. Thorough experiments show the superior performance of the baseline network, and the proposed CMCL can further bring performance over the baselines, outperforming the state-of-the-art methods considerably.