Person Re-identification by Saliency-Weighted Descriptor and Ranking Aggregation

Person re-identification which identifies the same person appeared in non-overlapping camera views is an important and challenging task in computer vision. Although most feature representation methods have significantly improved the person re-identification performance, they do not distinguish between pedestrian object and the environment in images in the process of extracting feature. In this paper, we present a novel feature representation called saliency-weighted descriptor (SWD) which intensifies the discrimination of pedestrian feature. Furthermore, we propose a ranking aggregation algorithm to combine SWD and unweighted descriptor for the purpose of mitigating the impact of inaccurate salient region. The experimental results on public person re-identification datasets (VIPeR, QMUL GRID, CUHK01, and CUHK03) demonstrate the effectiveness of our approach.