KRE: A Key-retained Random Erasing Method for Occluded Person Re-identification

Abstract

Occluded person re-identification (ReID) is a challenging task in the field of computer vision, facing the problem that the target pedestrians in probe images are obscured by various occlusions. Random Erasing in data augmentation techniques is one of the effective methods used to deal with the occlusion problem, but it may introduce noise into the training process, which affects the training of the model. In order to solve this problem, we propose an novel data augmentation method named Key-retained Random Erasing (KRE) which preserves the critical parts in images for occluded person ReID. Based on the regular Random Erasing, we utilize the naturally generated attention map in Vision Transformers and introduce an adaptive threshold selection method to detect the key areas of the image to be augmented. The complexity of the training samples can be improved without losing the key information of the images by reserving the key areas in Random Erasing process, which can finally alleviate the occluded person ReID problem. Validating the proposed method on occluded, partial and holistic ReID datasets, extensive experimental results demonstrate that our method performs favorably against state-of-the-art methods on ViT-based models.