Hierarchical Refined Local Associations for Robust Person Re-Identification

Abstract

Person re-identification is the process to identify a person from images/ videos captured from different nonoverlapping cameras in an autonomous way. The biological vision scheme emphasis on local discriminative cues in addition to the global appearance of a person for re-identification. The local cues are helpful to identify a person even if viewed at different scales and with different backgrounds. To emphasize on local cues, in this paper we present a refined association scheme for local parts of the images. The proposed scheme eliminates the effects of scale differences and background noise for automated person re-identification. Our approach divides the image of a person in horizontal strips and vertical sub-patches. A hierarchical refined associations based network (HRAN) is introduced to establish the refined associations among local segments of given images. In the first phase, the associations are established among horizontal strips of two images. In the next phase, the vertical sub-patches of associated horizontal strips are aligned/ linked with each other. Background noise and scale differences between images are addressed effectively using the proposed two-step mechanism. The triplet loss is used to optimize the refined local associations among images. A different weighting scheme is used for local and global losses for optimization of proposed model. The evaluation results of proposed methodology on two publicly available large scale datasets Market-1501 and DukeMTMC-ReID verified the effectiveness of proposed refined alignment method.