PA-Net: Learning local features using by pose attention for short-term person re-identification

Abstract

Person re-identification (Re-ID) is an important but challenging task in video surveillance applications. In Re-ID tasks, pose is an extremely useful cue to identify a person, even from the back view. Therefore, pose-detection models may learn the features that are beneficial to the Re-ID task and improve the Re-ID performance by fusing the feature maps into the Re-ID model. Two key problems in integrating the pose cues are addressed in this study. One is how to reduce the noise caused by cross-domain datasets. The other is how to fuse the feature maps to better utilize high-level semantic pose cues. To address these two key problems, we first propose PA-Net by combining the pose attention stream and the global attention stream, where the global attention stream distinguishes persons with different global appearances, and the pose attention stream distinguishes persons with similar global appearance but different poses. Then, we present a pose attention stream that learns local features to reduce the noise in the pose cues caused by the cross-domain datasets and provide more semantic information for the Re-ID task. The effects of the proposed pose attention are demonstrated in an ablation study, and comparative experiments show that PA-Net achieves state-of-the-art performance.