Fashion! Turn to the left! Computational model for visual stimuli
Aug 29, 2016
Clothing fashions are usually expressed with visual stimuli, for example style, color or texture. A collaboration of researchers from Shenzhen University, Sun Yat-sen University and University of South Carolina decided to use machine learning and computer vision techniques to see which visual stimulus has higher influence on updating of clothing fashion. They have proposed a classificatory model to quantify this influence by its corresponding accuracy in fashion classification.
With the development of computer vision algorithms, it is possible to parse visual stimuli of clothing into shape, color and texture patterns, which are understandable by machines, and therefore can be used for training, testing, and classification. Each visual stimulus use was studied as features to recognize the underlying fashion season of clothes: the more accurate the recognition, the more influential this visual stimulus is in clothing fashion. They have collected 9339 photos of fashioned clothes from 2008 to 2015 from 14 world-famous brands, and each year was treated as a different fashion season. Photos were categorized into 8 classes based on their season labels, and 8-class fashion classification was established using machine-learning algorithms and the according features were extracted and used for training and testing. Datasets that were used were STYLE130, COLOR335, Outex-TC-00021, and FASHION8, with different clothing styles and categories.
Three experiments were designed to select reliable feature descriptors for clothing style, color and texture description. Then, the analysis of clothing-fashion updates was mapped into a task of clothing-fashion classification and the selected descriptors were used. A ground-truth fashion dataset was constructed by collecting fashion photographs in continuous eight years based on the observation that the clothing fashion updates year by year. The experiments were conducted to evaluate the capability of style, color, and texture in clothing fashion classification.
The experimental results demonstrate that style holds a higher influence than the color, and that the color holds a higher influence than the texture.