Evaluation of basic visual features for person re-identification

Appearance-based person re-identification namely matching the same pedestrian across disjoint camera views, is an increasingly research spot. Researchers have proposed many complicated appearance descriptions. The role of basic visual features is seldom investigated, e.g. color, texture features, which are the basis of descriptions. By analyzing the characteristics of data environments, we can get the recommended basic features, while an improper feature selection may lead to suboptimal or even worse re-identification results. In this paper, an objective and comprehensive evaluation of the basic visual features is investigated and compared in the context of person re-identification. The performance of 13 representative color and texture features with two standard distance metrics are compared under two typical different environments. Basic feature recommendation for different environments was achieved by this evaluation. Furthermore, we verify the hierarchical structure of person re-identification network. Two state-of-the-art appearance descriptions are revised and improved by selecting more proper basic features. Extensive experiments show that basic feature evaluation is conducive to better feature selection.