Real-time Segmenting Human Portrait at Anywhere

Real-time portrait segmentation is an important task for a wide range of human-centered applications. With the increase of mobile devices, such as mobile phones and personal computers, more and more human-centered applications are transferred to running on these devices to provide users with a better experience. So, lightweight model designing becomes indispensable for building applications on these limited-resource platforms. In this work, we propose a real-time segmentation U-shape architecture with a Re-parameter Compress Residual module (RCR module) and a bypass branch that can further improve the segmentation efficiency. In order to speed up during the inference phase, the RCR module is compressed during inference, and the bypass branch adds the missing edge information improving the learning skill of the network. Based on the experiments on the EG1800 and P3M-10K dataset compared with the state-of-the-art methods, the proposed method achieves better performance with less number of parameters. Specifically, our method reduces the number of parameters around 50% while maintaining comparable high accuracy, and the details will be described in the experiment part.