An efficient feature pyramid attention network for person re-identification

Abstract

For person re-identification, occlusion, appearance similarity and background clutter have always been challenges. In order to effectively address the challenges, we propose an efficient feature pyramid attention network (FPA-Net), which combines visual features from different levels to focus on both detail features and information. Specifically, we embed a pair of attention mechanisms that complement each other in the backbone network to focus on the discriminant features of person areas. In addition, we designed a novel feature pyramid structure, which propagates the feature information from the cross-level through the top feature to the bottom feature and from the bottom feature to the top feature to supplement the detail information of the feature. Finally, we integrate features form different scales through a lightweight transition block to generate more discriminant features. Our method performed experimental analysis on four datasets: Market-1501, DukeMTMC-ReID, CUHK03(Detected) and MSMT17. A large number of experimental results prove that the performance of the method is significantly better than the existing state-of-the-art methods.