Clothes style recommendation system

We propose a clothes style recommendation system by analyzing the relation between facial features and clothes style. Five different kinds of face shapes and seven different kinds of clothes styles are defined in our work. To extract features that are stable under different lighting conditions, geometric information is used, which measure distance between regular features, e.g. distance between eyes, average distance from eye to nose. Instead of detecting regular facial features directly, facial feature points are detected by active shape model in advance. Then 14 different kinds of geometric information are extracted, which can capture discriminant features to describe the significance properties not only for the specific facial shape but between different facial shapes. Finally, multi-label classification is applied because one facial shape is suitable to more one clothes styles. Binary-Relevance (BP) and Label Powerset (LP) methods are used to transfer multi-label classification into multiple binary class problems and one multi-class problem, respectively. Experiments are designed to evaluate the system performance with two transferring methods, and Hamming-loss function and F-score are used for accuracy measure.