Augmentation to Disentangling: A Novel Approach for Improving Feature Disentanglement in Occluded Person Re-Identification

Occluded person re-identification poses a significant challenge due to the obstruction of human body parts by various obstacles, requiring the use of incomplete or invisible body parts for matching. Existing methods often address occlusion through feature disentanglement. However, the absence of occlusion labels hampers accurate disentanglement, leading to challenges in precisely distinguishing occluded regions. In response to these challenges, we propose a novel approach that actively introduces occlusion to guide feature disentanglement. Our method incorporates an Occlusion Sample Augmentation module, actively introducing occlusion to person images and generating accurate occlusion labels. Additionally, we employ pose estimation to partition the human body into distinct semantic components, enabling selective matching of non-occluded parts. To address interference from occlusion, we introduce an Occlusion-augmented Push Loss. This loss function effectively separates human body features from occluded part features based on augmented occlusion location labels, guiding the learning of feature disentanglement. Experimental results on both the Occluded-Duke and DukeMTMC-reID datasets, encompassing occluded and holistic re-identification tasks, showcase the outstanding performance of our method, achieving state-of-the-art (SOTA) performance.