Person re-identification via learning visual similarity on corresponding patch pairs

Abstract

Since humans concentrate more on differences between relatively small but salient body regions in matching person across disjoint camera views, we propose these differences to be the most significant character in person re-identification (Re-ID). Unlike existing methods focusing on learning discriminative features to adapt viewpoint variation using global visual similarity, we propose a learning visual similarity algorithm via corresponding patch pairs (CPPs) for person Re-ID. The novel CPPs method is introduced to represent the corresponding body patches of the same person in different images with good robustness to body pose, viewpoint and illumination variations. The similarity between two people is measured by an improved bi-directional weight mechanism with a TF-IDF like patches weight. At last, a complementary similarity measure and a mutually-exclusive regulation are presented to enhance the performance of Re-ID. With quantitative evaluation on public datasets, the best rank-1 matching rate on the VIPeR dataset is improved by 4.14%.