SNAFA-Net: Squared Normalization Attention and Feature Alignment for Visible-Infrared Person Re-Identification

Jiahao Wu, Qiang Chen, Guoqiang Xiao

Abstract: Visible-infrared person re-identification (VI-ReID) aims to match images of the same individual across visible and infrared image sets. Due to the modality discrepancy between different modalities and the inter-modality discrepancy caused by occlusions and postures, make this a challenge task. In this paper, we propose a novel network, called SNAFA-Net, that uses innovative techniques to address the key challenges in VI-ReID. The squared normalization attention (SNA) module addresses the modality disparity by attending to the invariant features of pedestrians in each modality. The feature alignment-enriched part-based convolutional block (FA-PCB) module allows us to partition the human body image into multiple parts and extract corresponding features for each part of the body. These features are then aligned using the shortest path method, reducing the impact of posture and viewpoint changes. This alignment process ultimately reduces feature redundancy and improves the feature representation and discrimination accuracy. The experiments performed on the SYSU-MM01 [18] and RegDB [14] datasets demonstrate that the proposed approach achieves state-of-the-art performance.