Clothing Co-Parsing by Joint Image Segmentation and Labeling

摘要

This paper aims at developing an integrated system of clothing co-parsing, inorder to jointly parse a set of clothing images (unsegmented but annotated withtags) into semantic configurations. We propose a data-driven frameworkconsisting of two phases of inference. The first phase, referred as "imageco-segmentation", iterates to extract consistent regions on images and jointlyrefines the regions over all images by employing the exemplar-SVM (E-SVM)technique [23]. In the second phase (i.e. "region co-labeling"), we construct amulti-image graphical model by taking the segmented regions as vertices, andincorporate several contexts of clothing configuration (e.g., item location andmutual interactions). The joint label assignment can be solved using theefficient Graph Cuts algorithm. In addition to evaluate our framework on theFashionista dataset [30], we construct a dataset called CCP consisting of 2098high-resolution street fashion photos to demonstrate the performance of oursystem. We achieve 90.29% / 88.23% segmentation accuracy and 65.52% / 63.89%recognition rate on the Fashionista and the CCP datasets, respectively, whichare superior compared with state-of-the-art methods.