Automatic Matting Using Depth and Adaptive Trimap

Ying Zhao, Chao Chen, Liyan Liu, Wei Wang

Abstract: Image matting refers to the problem of accurately extracting the foreground object from an image. The trimap containing labels of known foreground, known background and unknown has been broadly used to reduce solution space of the problem. The matte of an unknown pixel can be solved using samples from its neighbor known foreground and known background. However, the existing methods of color-based sampling may fail when foreground and background share similar colors, meanwhile, automatically generated trimap may not properly cover foreground boundary. In this paper, we propose a novel matting method using depth-assisted sampling and adaptive trimap generation. We use depth to assist color for improving sample selection and generate trimap based on color distribution of local unknown regions to make it cover foreground boundary adaptively. Our experiments show the effectiveness of proposed method.