Selective relation-aware representations for person re-identification

Xi Luo, Min Jiang, Jun Kong

Abstract: Recent studies have shown the effectiveness of the joint learning of global and part-level features for person re-identification (Re-ID) task, as they target to enhance the discriminative feature representations of pedestrian from the perspective of multiple granularities. However, most of these methods ignore the global structural relation information under adaptive receptive fields, and the local structural relation between parts, making it hard to fully utilize rich structural relation information of pedestrian to form the stable coherent structural patterns. In this paper, we propose selective relation-aware representations (SRAR) for person Re-ID. The framework of SRAR mainly consists of a distinctive attention module and a part relation-aware (PRA) module. Firstly, the distinctive attention module embraces selective position attention module and channel attention (CA) module, which are designed to selectively exploit spatial-wise and channel-wise relation-aware salient feature representations under adaptive receptive fields for the pedestrian. Furthermore, to comprehensively grasp the coherent structural patterns of pedestrian, the PRA module is introduced to promote the interaction between part-level features, and further obtain relation-aware fine-grained feature representations. Extensive experiments are performed on the three popular datasets, including Market1501, DukeMTMC-reID, CUHK03, which validate that our method achieves a competitive performance.