Enhancing Person Re-identification Across Modalities with Local and Global Feature Correlation

We investigate the challenging problem of visible infrared person re-identification (VI-ReID). VI-ReID addresses pedestrian retrieval by matching images across different modalities. Given a infrared(visible) image of a person, the VI-ReID module finds the corresponding images from the visible(infrared) gallery set. This task is particularly difficult due to significant differences between visible and infrared images, such as color and texture. For VI-ReID, learning discriminative features from cross-modal images is crucial. In this paper, we propose a novel approach for modeling local and global structural relationships and capturing both modality-specific and modality-shareable information. Specifically, we introduce the Global Feature Attention Module (GFAM) to learn the feature correlation between two modalities and the Local Feature Alignment Module (LFAM) to align intra-modality feature representations. Our end-to-end unsupervised learning architecture also employs the InfoNCE loss to maximize the mutual information between positive samples while minimizing the mutual information between negative samples. Experimental results conducted on two benchmark datasets demonstrate that our proposed method outperforms state-of-the-art methods for both all-search and indoor-search situations, achieving gains of 2.05% Rank-1 and 1.12% mAP, respectively.