Pose-Guided Occlusion-Aware Network for Occluded Person Re-Identification

In real-world applications, person re-identification (ReID) often faces varying degrees of occlusion, resulting in missing target features and feature confusion caused by occluding objects. Existing methods using human body cues can primarily mitigate occlusion caused by objects. However, they often fail to distinguish between multiple people, making it difficult to resolve ambiguities caused by non-target pedestrians occluding the target. To address this problem, we propose PONet, which provides a comprehensive solution to occlusion robustness. Specifically, we design the POC-SE module, which explicitly simulates occlusion scenarios involving different objects and people and extracts discriminative features from key regions of the target pedestrian by exploiting joint pose constraints and occlusion prediction mechanisms. We adopt the Swin Transformer with cross-scale attention as the encoder, which enhances the model's focus on key semantic parts while improving high-resolution feature representation. In addition, we propose the prototypical contrastive loss, which mitigates the intra-class variance problem caused by occlusion within samples by integrating the intra-class distribution. Experimental results demonstrate that PONet achieves advanced performance in occlusion scenarios and holistic ReID benchmark tests, fully validating its effectiveness and robustness. Especially on the Occluded-Duke dataset, our method achieves 71.3% mAP and 82.9% Rank-1 accuracy.