DEEP GLOBAL AND LOCAL SALIENCY LEARNING WITH NEW RE-RANKING FOR PERSON RE-IDENTIFICATION

Feature representation and similarity metric are two key issues in person re-identification (re-id). In conventional feature extraction, spatial patches of persons are processed indiscriminately, while the uniqueness and difference of local patches are often ignored. In addition, re-ranking is not be paid sufficient attention. In this paper, a re-id algorithm based on a novel saliency learning and re-ranking is proposed to address above problems. Specifically, a new saliency learning method based on a three-stream CNN is first presented to learn distinctive features for upper-body, lower-body and global body. Besides, by modeling the interrelationships between the query and the gallery images, a reweighting-based re-ranking scheme is designed to improve the initial metric matrix. Experimental results on three public datasets, i.e., CUHK03, CUHK01 and VIPeR, indicate that the proposed method achieves very competitive performance compared with the state-of-the-art approaches.