Single-shot person re-identification by spectral matching of symmetry-driven local features

Abstract

This paper presents an appearance-based approach for person re-identification. It comprises the extraction of features that models the significant prospects of human appearance: local patterns of the regions and the presence of the spatial distribution of color, texture with high informative patches. This information is extracted from different body parts by following the symmetry-asymmetry precept and combined to form the signature of an individual. In this way, the signatures provide robustness against illumination variation, arbitrary pose changes, and occlusion. Further, these signatures are utilized for the process of re-identification through the spectral matching scheme. The approach applies to a situation where single image per individual is available. Extensive experiments are carried out on challenging datasets to validate the performance of proposed model over the stateof-the-art.