Multi-modal deep network for RGB-D segmentation of clothes

In this Letter, the authors propose a deep learning based method to perform semantic segmentation of clothes from RGB-D images of people. First, they present a synthetic dataset containing more than 50,000 RGB-D samples of characters in different clothing styles, featuring various poses and environments for a total of nine semantic classes. The proposed data generation pipeline allows for fast production of RGB, depth images and ground-truth label maps. Secondly, a novel multi-modal encoder-ecoder convolutional network is proposed which operates on RGB and depth modalities. Multi-modal features are merged using trained fusion modules which use multi-scale atrous convolutions in the fusion process. The method is numerically evaluated on synthetic data and visually assessed on real-world data. The experiments demonstrate the efficiency of the proposed model over existing methods.