Infrared-visible person re-identification via Dual-Channel attention mechanism

Infrared-Visible person re-identification (IV-ReID) is really a challenging task, which aims to match pedestrian images captured by visible and thermal cameras. There exists differences in appearance between visible and infrared images caused by viewpoint changes, pose variations and deformations, and additional cross-modality gap caused by different camera spectrums. These discrepancy make IV-ReID difficult to be addressed. In order to solve this problem, we propose a dual-path network with an attention mechanism called Convolutional Block Attention Module (CBAM) to learn the discriminative feature representations, and a modified Batch Norm Neck (BNNeck) module fuses the feature representation of cross-modality to improve the identity recognition accuracy. Specifically, the proposed method firstly constructs two independent networks to learn modality-specific feature representation, next the feature representation is split into several stripes by a conventional average pooling layer, then a shared layer is introduced to project feature representation from cross-modality into the same embedding space. Finally, we fuse heterogeneous loss function and cross entropy loss function to measure the feature similarity to improve the performance. The experimental results on two public cross-modality person re-identification datasets (SYSU-MM01 and RegDB) demonstrate that the proposed method can significantly improve the performance of IV-ReID.