Pose-Guided Feature Disentangling for Occluded Person Re-identification Based on Transformer

Occluded person re-identification is a challenging task as human body parts could be occluded by some obstacles (e.g. trees, cars, and pedestrians) in certain scenes. Some existing pose-guided methods solve this problem aligning according to graph matching, but these graph-based are not intuitive complicated. Therefore, we propose transformer-based Pose-guided Feature Disentangling (PFD) method utilizing pose information clearly disentangle semantic components or joint parts) selectively match non-occluded correspondingly. First, Vision Transformer (ViT) used extract the patch features with its strong capability. Second, preliminarily from information, matching distributing mechanism leveraged Aggregation (PFA) module. Third, set of learnable views introduced transformer decoder implicitly enhance disentangled part features. However, those guaranteed related without additional supervision. Pose-View Matching (PVM) module proposed explicitly visible automatically separate occlusion Fourth, better prevent interference occlusions, design Push Loss emphasize parts. Extensive experiments over five datasets for two tasks (occluded holistic Re-ID) demonstrate that our PFD superior promising, which performs favorably against state-of-the-art methods. Code available at https://github.com/WangTaoAs/PFD_Net