Visible-Infrared Cross-Modal Pedestrian Re-identification based on Two Attention-calibrated Correlation Features

Abstract

Visible-Infrared Cross-Modal Pedestrian Re-identification faces the challenges of feature misalignment due to viewpoint differences and pose changes, and insufficient robustness to noisy samples. To this end, this paper proposes a cross-modal pedestrian re-identification based on two attention calibrated associative features (Two Attention Calibrated Associative Feature Networks, TACAFNet). Firstly, a data enhancement strategy based on channel relationship is designed to use to generate diverse samples through uniform sampling of channel perception and simulate occlusion scenarios by combining with a random channel erasure technique to improve the model's generalisation ability to cross-modal colour differences. Secondly, the Image Pair Correlation Attention Module (IPCAM) is designed to construct graph structural relationships by exploiting the contextual relevance of intra-modal pedestrian features, to enhance feature discriminative properties and to suppress background noise interference. Further, Cross-modal Alignment Attention Module (CMAAM) is proposed to reduce modal correlation interference through inter-modal component-level feature matching and enhance cross-modal fine-grained feature alignment to reduce intra-class differences. Experiments on SYSU-MM01 and RegDB datasets show that TACAFNet significantly outperforms existing mainstream methods, validating the effectiveness of the proposed model in the cross-modal pedestrian re-identification task.