Abstract
We proposed a compact color descriptor specialized in expressing clothing for person re-identification across cameras. It expresses any type of clothing colors by combination of a set of unicolor clothing selected from a collection of various colored clothing widely and densely spread over clothing color space; which is called a wardrobe. Proper wardrobe can be collected through a clothing manufacture having a large variety of colors. Effective selection of proper clothing from the wardrobe is performed by using the coefficients learned by linear SVM with L1 regulation. In the evaluations on the highly challenging VIPeR dataset, wider coverage of color space of our wardrobe is proved by comparing 11 grouped colors with the state-of-the-art descriptor; color names of the same dimension. And the proposed clothing selection by SVM with L1 regulation shows its effectiveness by achieving improvement on accuracy while reducing dimension.