Composite Augmentation and Feature Integration Reconstruction for Occluded Person Re-Identification

Most existing occluded person re-identification (Re-ID) methods rely on external cues to separate the visible parts of the person from the occluded regions, aiming to achieve alignment based on visible body parts. However, such methods often overlook the complexity of different occlusion scenarios. Moreover, there exists a severe imbalance between the training and testing sets in current occlusion datasets, which further undermines the model's generalization ability. In this paper, we propose a method named Composite Augmentation and Feature Integration Reconstruction For Occluded Person Re-Identification. Firstly, a Composite Data Augmentation Module is designed to increase the number of occluded samples in the training set, thus alleviating data imbalance. Next, a feature interaction module is introduced to perform bidirectional interactions between global and local features, which helps reduce the impact of occlusion-induced noise and feature redundancy in complex scenes, thereby improving feature representation efficiency. Finally, a Feature Reconstruction Module retrieves the top-k nearest neighbor images from the gallery to reconstruct the occluded body parts, enabling more complete person matching. Experimental results on five challenging occlusion datasets demonstrate that the proposed method achieves superior performance.