Pose-Aware Feature Selection Enhanced Network for Person Re-Identification

Achieving robust person re-identification remains an open research problem in visual surveillance. Most previous models have made significant progress in extracting Person features, but they often pay less attention to local features that carry additional clues. To tackle this problem, in this paperwe propose a pose-aware feature selection enhancement network. Firstly, we designed a multi-scale bidirectional expansion fusion module. In this module, each pixel point of the high-level semantic feature map is expanded along the horizontal and vertical directions based on its position, aligning and fusing high-level features with low-level features of the same size to extract multi-scale features of Person samples. Secondly, we introduced a multi-channel feature selection enhancement module, which divides the human body pose key point features and the fused feature areas, allowing all channels to compare differences in a set manner, selecting the features that need to be enhanced, and then splicing the enhanced features with the fused features, enabling the model to simultaneously learn Person features and local features with additional clues. We conducted extensive experiments and analyses, and through experiments on a widely used dataset, we have demonstrated the superiority of our method compared to many state-of-the-art methods.