Real-time multi-view background matting for 3D light field video

Matting is a method to extract foreground objects of arbitrary shape from an image. In the field of 3D display, matting technology is of great significance. Through the study of this technology, we can extract high-quality target foreground, and then reduce unnecessary stereo matching calculation and improve the effect of 3D display. This paper primarily studies the human target in 3D light field, and proposes a real-time multi-view background matting algorithm based on deep learning. Three-dimensional video live broadcast puts forward high requirements for the real-time performance of the matting algorithm. We pre-compose a group of multi-view images taken at the same time into a multi-view combined image. The network directly carries on the background matting to the multi-view combined image and outputs a group of foreground images at one time. Because the background of the multi-view combined image is not holistic, a pre-photographed background picture without human is added to the input to assist the network for learning. In addition, we add a channel subtraction module to help the network better understand the role of the original image and background image in the matting task. The method in this paper is tested on our multi-view data set. For pictures with different background complexity, it can run about 65 frames per second and maintain a relatively stable accuracy. The method can efficiently generate multi-view matting results and meet the requirements of 3D video live broadcast.