A New Discriminative Feature Learning for Person Re-Identification Using Additive Angular Margin Softmax Loss

Abstract

In this paper, a new end-to-end framework is proposed for person re-identification (re-ID) by combining metric learning and classification. In this new framework, the Additive Angular Margin Softmax is used which imposes an additive angular margin constraint to the target logit on hypersphere manifold. This is aimed to improve the similarity of the intra-class features and the dissimilarity of the inter-class features simultaneously. Compard with the three popular used softmax-based-loss methods, the experiments show that the proposed approach has achieved improved performance on Market1501 and DukeMTMC-reID datasets for person re-ID.