LEARNING DISCRIMINATIVE FEATURES THROUGH AN INDIVIDUAL'S ENTIRE BODY AND THE VISUAL ATTENTIONAL PARTS FOR PERSON RE-IDENTIFICATION

Person Re-Identification (Re-ID) aims to match a specific person across different camera views, which has wide application in public security and image retrieval. For example, Re-ID can help the police get trajectories of suspects. Re-ID still remains a challenging task due to large variations in illumination, background clutter, occlusion and human pose. In this work, a novel deep learning architecture containing global and attentional branches is proposed to learn discriminative representations of persons in differing contexts for Re-ID. Specifically, the global branch is a traditional deep model that learns global features with the images of a person. The attentional branch uses a low-rank approximation of a bilinear pooling model to learn attentional maps by automatically focusing on the visual attentional parts of an individual. The whole model is trained jointly in an end-to-end method. The features of the entire body and visual attentional parts obtained by the trained model are concatenated as representations of persons. Finally, a generic cosine distance metric is used for the person Re-ID task. Extensive experiments on several benchmark datasets including CUHK01, CUHK03 and Market1501 demonstrate the effectiveness of our method compared to the current state-of-the-art approaches.