Valid Keypoint Augmentation based Occluded Person Re-Identification

Abstract

Occluded person re-identification is a challenging task which aims to search for or distinguish the specific person as human body is occluded by obstacles or other persons or by oneself. Some recent State of the art works which adopt transformer and/or pose-guided methods improve the feature representation and performances, but there is a room to enhance them in both representation and heavy structure. In this paper, we suggest to efficiently improve the transformer-based Re-ID method for the occluded person as follows. First, in data augmentation to improve Re-identification performance, instead of deleting an arbitrary area, only the part containing the keypoint feature of a person is deleted for effective learning in occlusion. Second, a consistency loss between global and local features of a body part is proposed for improving the discrimination to recognize the identical person. We compare the mAP and Rank-1 performances of our approach and various existing methods on the Occluded-Duke dataset. Experimental results show that our proposed model outperforms the competitive methods.