Clothing Co-Segmentation Based on HOG Features and E-SVM Classifier

To improve the accuracy of clothing segmentation, a novel co-segmentation method is proposed in this paper based on HOG(Histogram of Oriented Gradients) features and E-SVM(Exemplar Support Vector Machine) classifier. The cosegmentation method uses an auxiliary dataset and is implemented with three steps: superpixel grouping, E-SVM classifier training, and segmentation propagation. Firstly, image input by user is segmented in superpixel, such that the images in auxiliary dataset can be divided into multi-regions. Secondly, some regions are selected to a positive segmentation and HOG information is used to train E-SVM classifiers. Finally, the cloth is segmented into user image with E-SVM classifiers and segmentation propagation. Experimental results show that the proposed method can segment clothing images with high accuracy.