Conditional Feature Coupling Network for Multi-persons Clothing Parsing

Abstract

Clothing parsing provides some significant cues to analyze the dressing collocation and occasion. In this paper, we propose a novel clothing parsing framework with deep end-to-end conditional feature coupling network for the photographic multi-persons in the fashion scene, and annotate a multi-persons clothing dataset for the effectiveness demonstration. Our parsing framework has three sub-networks, including the coarse parsing network (CPN), the multi-pose feature network (MFN) and the coupling residual network (CRN). CPN and MFN generate a coarse segmentation intermediary and 28 pose-indicated heat maps, respectively. CRN receives these auxiliary information and generates the fine-tuning clothing parsing result. To verify the generality and effectiveness of our parsing framework, we compare our method with the state-of-the-art parsing and segmentation methods such as Deeplab [2] and Co-CNN [7] on our multi-persons clothing dataset and some fashion clothing benchmarks. Experimental evaluations on these datasets demonstrate that our framework has a superior performance in the parsing task. In particular, our CFCN achieves 88.74% accuracy on the multi-persons clothing dataset, which is significantly higher than 86.50% by Deeplab. The project is available at https://github.com/suzhuoi/CFCNet.

This research is supported by the National Natural Science Foundation of China (61502541, 61772140, 61502546), the Natural Science Foundation of Guangdong Province (2016A030310202), the Science and Technology Planning Project of Zhongshan (2016A1044), and the Fundamental Research Funds for the Central Universities (Sun Yat-sen University, 16lgpy39).