Joining features by global guidance with bi-relevance trihard loss for person re-identification

Person re-identification (ReID) aims to associate the person with the given identity across different cameras, which has wide application in the field of intelligent video. In this work, an efficient method is proposed to improve ReID performance. First, a global-guided feature joint network is designed, which consists of a multiple feature extraction network and a global-guided feature fusion network. The former can align different body regions and extract global feature and local features. The latter utilizes the global feature to guide the adaptive fusion of the local features, which can evaluate the importance of different local features dynamically. Secondly, a Bi-relevance TriHard loss (TriBR) is designed to penalize the loss dynamically. TriBR combines Euclidean distance and angle information, which considers the self-relevance of the intra-class samples and the cross-relevance of the inter-class samples simultaneously. Besides, TriBR can adaptively adjust the distance margin and the angle margin to optimize the network. The proposed TriBR is advantageous to learn more discriminative features. The method achieves 89.2% mAP (mean Average Precision) with 95.8% Rank-1 on Market-1501, 79.7% mAP with 89.3% Rank-1 on DukeMTMC. The proposed method also performs excellently on the datasets for the occluded person re-identification problem.