Person Re-Identification via Locally Biased Metric Learning

Abstract

Person re-identification, which means matching person across non-overlapping cameras in a surveillance camera network, has attracted more and more attention. A lot of metric learning based methods, which generally learn a new distance function under two pair-wise constrains, i.e. similar constrain and dissimilar constrain, were proposed to address the challenging problem due to significant appearance variances caused by pose changes, lighting variations and image resolution differences. However, these methods attempt to satisfy all similar constrains and dissimilar constrains, which may be conflict and cannot be simultaneously satisfied in the practical application. In this paper, we propose a new local metric learning method based KISS metric learning. Comparative experiments conducted on three public standard datasets have shown the promising prospect of the proposed method.