Multi-resolution feature perception network for UAV person re-identification

Person re-identification (re-id) with unmanned aerial vehicles (UAVs) is of great significance in intelligent surveillance. However, recognizing a person of interest from UAVs is more challenging than existing person re-id tasks across multiple fixed cameras. The images taken by UAVs have large resolution variations and complex backgrounds due to the rapid movement and constantly changing flight altitudes of UAVs. Some methods propose cross-resolution learning for person images captured by fixed cameras, assuming query images with low resolution (LR) and gallery images with high resolution (HR). However, they are incapable of handing the resolution variations in UAV scenarios, where both query and gallery images are with significant resolution variations. In this paper, we present a novel multi-resolution feature perception network (MRFPN) to learn discriminative and resolution-robust feature for UAV person re-id. Firstly, we introduce a self-attention module to capture the full-image context information in pixel level and obtain the pixel context-aware feature map for both HR and LR images, which can effectively deal with the background clutters. Secondly, we construct a cross-attention module to learn resolution-robust representations by bi-directionally perceiving the resolution-guided semantic information between HR and LR features. Further, we design a semantic consistency constraint to limit the difference of HR and LR features. Extensive experiments show the superiority of our method on both UAV and traditional datasets.