Optimizing multi-granularity region similarity for person re-identification

Abstract

Person re-identification (re-id) is one of the hottest research topics due to its great value in video analysis applications, such as indoor security and road surveillance. It has been verified as beneficial for re-id to joint global and local features in the recent literature. However, most existing methods usually extract the features of the global region or divide the whole image into several parts without considering the alignment of different parts, which are not discriminating or robust to the complex scenarios. In this paper, we propose a novel method that optimizes multi-granularity similarity fusion based on the coarse region and the fine region. We extract the global feature representations of the coarse region and the local feature representations of the fine region. Instead of using the pose estimation method, we align the local parts by calculating the similarity between the local parts, which is optimized by the refined longest path. The extensive experiments on four challenging datasets are carried out and indicate that our method has achieved the state-of-the-art performances. For instance, on the VIPeR dataset, our method achieves a 67.25% rank-1 matching rate, outperforming the state-of-the-art approaches by a large margin.