Occluded Person Re-Identification via Multi-Branch Interaction.

Person re-identification (re-ID) aims to retrieve images of a given individual from different camera views. Obstacles obstructing parts of a pedestrian's body often result in incomplete identity information, impairing recognition performance. To address the occlusion problem, a method called Multi-Branch Interaction Network (MBIN) is proposed, which exploits the information interaction between different branches to effectively characterize occluded pedestrians for person re-ID. The method consists primarily of a hard branch, a soft branch, and a view branch. The hard branch enhances feature robustness via a unified horizontal partitioning strategy. The soft branch improves the high-level feature representation via multi-head attention. The view branch fuses multi-view feature maps to form a comprehensive representation via a dual-classifier fusion mechanism. Moreover, a mutual knowledge distillation strategy is employed to promote knowledge exchange among the three branches. Extensive experiments are conducted on widely used person re-ID datasets to validate the effectiveness of our method.