| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 100 === This study presents a novel approach to automatically detect clothing genre from a full-body input image with no restrictions of people poses, races, genders, ages, image backgrounds, and image resolutions. We have identified five style elements for distinguishing eight kinds of upperwear genres and seven style elements for distinguishing eight kinds of lowerwear genres. These style elements are identified based on the clothing design theory. The corresponding features for each style elements are also designed. For extracting these style elements, in the proposed frameworks, we first employ an upper-body detector to get body parts of a person standing arbitrary pose in images. The extracted feature vectors of the style elements are then exploited for learning the genre models and used for predicting the clothing genre. We illustrated the effectiveness of our approach by showing the achieved precision, recall, and F-score. For upperwear genres detection, we achieved overall precision of 92.04%, recall of 92.45%, and F-score of 92.25%. And for lowerwear genres detection, we achieved overall precision of 94.66%, recall of 93.64%, and F-score of 94.15%.
|
|---|
