HPAN: A Hybrid Pose Attention Network for Person Re-Identification

To address the difficulty in expressing the correlation between different local features extracted by the current person re-identification feature extraction methods, and the challenge of effectively integrating local features with global features, a Hybrid Pose Attention Network (HPAN) for person re-identification is proposed. In HPAN, the high-resolution network HRNet-W32 serves as the backbone for person re-identification and pose estimation, extracting global features and local key point heatmaps of the human images, and then generating local key point features. Self-attention is used to extract the correlation between each local key point feature, generating local pose features. Furthermore, a Hybrid Pose and Global Feature Fusion (HPGFF) module is adopted to fuse the global features and local pose features, creating integrated features. To evaluate, we conduct experiments on five publicly available datasets, and HPAN has all achieved competitive or state-of-the-art results.