PGAN: A Position-based Graph Attention Network for Visible-infrared Person Re-identification

Matching persons in various environment of infrared and visible images is a retrieval task known as visible-infrared person re-identification (VIReID). The primary obstacle of VIReID is the modality gap between visible and infrared images. To this end, we proposed a novel dual-stream model named Position-based Graph Attention Network (PGAN) to reduce the gap. It contained with the Channel Spatial Local Attention (CSLA) module and Graph Attention Transform (GAT) module. Channel Spatial Local Attention module generated 7 blocks of learnable position encoding. And Graph Attention Transform module evolved from ResNet50 with graph attention mechanism. In the process of training, we designed a suitable loss function to constrain the generation of position encoding. Extensive experiments were implemented on SYSU-MM01 and RegDB datasets, effectively demonstrating the superiority of proposed PGAN over other SOTA algorithms.