Analysis of Visual Attributes of Clothes on Dressing Images

We propose a new method to extract the attributes of clothes on human body. Some appropriate composite-parts of the collar are chosen to represent its peculiarity. The HOG feature is extracted and efficient filters with linear SVM are created to predicate the parts. Then structure expanding method is proposed to recognize the collar type according to the constraint spatial relationships between the parts. We also acquire the clothes length attribute with the area search method and analyze the sleeve length attribute. A contrast experiment on a dataset including 1683 images indicates that our methods can significantly improve the recognition accuracy. Finally we develop a fully-automatic image retrieving system to show the effect of our work.