Tagging the shoe images by semantic attributes

With the rapid proliferation of Internet, it becomes a great challenge to annotate explosive number of objects manually. Especially for the fashion domain where a massive collection of new products come up everyday. Therefore, to save human labor, it is essential to develop an automatic tagging system for those fashion products in a variety of appearances. In this paper, we focus on addressing the issue of automatic shoe tagging where a novel system is proposed to predict the semantic attributes of the shoe images. Given a shoe image in an unknown viewpoint, our proposed system first classify it into one of the 6 pre-defined representative viewpoints, which are commonly displayed in online merchants. To localize the shoe parts on the identified viewpoint, a view-specific part localization model is proposed based on the prior knowledge of the shoe structures under different viewpoints. Finally, we extract several complementary low-level features from the localized shoe parts, which is fed into a SVM classifier for attribute prediction. The effectiveness of the proposed system is demonstrated on a newly-built pump shoe image dataset.