Global-to-Local Feature Mining Network for RGB-Infrared Person Re-Identification

RGB-Infrared person Re-Identification (RGB-IR ReID) is a challenging matching task that retrieves a RGB/infrared pedestrian image from the existing infrared/RGB set captured by non-overlapping visible or infrared cameras. Existing works mainly focus on how to alleviate the intra-modality variations and inter-modality discrepancies by data augmentation or feature alignment. Although these methods enlarge the diversity of the training set and, to some extent, reduce the gap between modalities, insufficient mining of discriminative and invariant features between modalities limits the performance of RGB-IR ReID algorithms. To remedy this, we propose a global-to-local feature mining network (GFMNet) to further mine discriminative and invariant features. Specifically, GFMNet contains two feature mining modules: Attention-aware Feature Mining Module (AFMM) and Local Information Mining Module (LIMM). AFMM aims to learn global discriminative features by attention mechanism; LIMM mines potential local invariant features between modalities by shortest path exploration. Besides, to reduce modality discrepancies and define a unified convergence direction, we introduce distribution consistency (DC) loss, which encourages RGB and infrared modalities toward intermediate modality. Extensive experiments on the SYSU-MM01 and RegDB datasets show that GFMNet achieves competitive RGB-IR ReID performance. The code will be announced at https://github.com/cq0907/GFMNet .