Covariance based person re-identification using spectral matching

This paper presents an approach for reidentification based on appearance. The person re-identification is recently introduced and yet an unsolved problem in computer vision. Re-identification refers to identify an individual who has already been observed by different cameras. The appearance of an individual in different cameras looks unlike due to illumination variations and arbitrary pose alternations. The identity of an individual is represented by a distinct signature model that should invariant to illumination, pose variation and occlusions. This paper focuses on the formation of distinct signature models based on mean covariance patch. A patch homogeneity is proposed which handles the clutter in the image of a specific individual. The signature model of each individual needs to find its corresponding signature model over the network. The idea of spectral matching is used for the computation of matching between the models signature. The matching signature models are ranked according to matching scores. The performance of our approaches is evaluated on ETHZ and VIPeR data sets and the results are shown in cumulative matching characteristics.