Visible-Infrared Person Re-Identification Based on Dual Branch Structure and Attention

Visible-infrared person re-identification (VI-ReID) is a difficult cross-modal pedestrian retrieval task.due to the modality gaps between visible (VIS) and infrared (IR) images. In the process of visible-infrared person re-identification, it is difficult to extract effective information from modality-shared features due to the additional cross-modality discrepancy. To address this problem, we propose a visible-infrared person re-identification method based on dual-branch structure and attention. Firstly, a dual-branch convolution module is added to the ResNet-50 backbone network to extract more plentiful modality-shared features of different scales by using the dual-branch structure and dilated convolution with different expansion rates. Moreover, we design a Non- Local attention module to obtain the correlation between different pixels in spatial and channel dimensions by introducing and improving the Non-Local blocks, thereby enhancing the ability of the network to extract structural information. Finally, we design an efficient joint loss strategy to optimize the network to further reduce the modality discrepancy. Our method shows excellent performances on both SYSU-MM01 and RegDB datasets.