Exploiting color cues to improve person re-identification

Person re-identification is essentially the task of recognizing the same person across different non-overlapping cameras. It has been intensively studied due to its critical role for many security applications of video surveillance. In this paper we exploit the low-level color information in two different aspects, showing the strength provided by colors to increase the accuracy of classification. First, we propose a discriminant embedding for the feature descriptor, which takes into consideration all color components at one time using the quaternions. Second, we propose to assign each individual to a color name in order to increase the discrimination of our descriptor with a semantic analysis. Experiments are carried out on the highly challenging VIPeR dataset. Comparison with some state-of-the-art methods is provided and the proposed method shows better performance even with a simple metric learning method.