Discriminative Transfer Learning Siamese CNN for Person Re-identification

Abstract

Person re-identification (Re-ID) has become an increasingly popular computer vision problem. It remains challenging, especially when there are non-overlapping cameras. In this paper, we review the two representative architecture, i.e., identification and verification models. They both have their advancements and limitations. We present a novel method to address the Re-ID problem. First, combine the two models to consist a more effective fusion loss function. Second, we find that CNNs which are pre-trained on large image datasets learn more discriminative knowledge with objective semantic, which can be transferred to subsequent layers to promote accuracy significantly. Experiments on four benchmark datasets show the superiority of our method over the state-of-the-art alternatives.