Person re-identification is the problem of matching pedestrian images observed by different cameras in non-overlapping regions. Semantic features, also called attributes, have demonstrated to produce state-of-the-art performances in this problem. In existing works, attributes are detected independently to each other. In this paper, we propose using relationships between attributes to refine the attribute detection result. Experimental results on two datasets VIPeR and PRID prove the effectiveness on performances when our method is applied.
展开▼
