GPU Accelerated Video Matting

Matting is a main component of the image and video processor, which has been widely used in video effect design. Along with the development of the information technology and the growth of people’s appreciation, high-definition and high resolution contents have become the mainstream of everyday life. Consequently, video matting technology faces the challenge of growing data sizes and increasing high demands on efficiency. In this paper, we present an robust video matting system with CPU, and examine the efficiency of main steps under variety inputs. And we then accelerate the two bottleneck steps of the system, SIFT and optical flow, using GPUs, whose parallel processing power has been developing rapidly in recent years, making it an ideal solution for the scalability issues of video matting. We use BSGP, a high-level general purpose GPU programming language, to accelerate the SIFT algorithm and optical flow algorithm of the video matting system. Experiments show that the overall efficiency of the system has been enhanced more than two times, and the acceleration is more obvious for higher resolution videos, making it a promising solution for high resolution video matting.