Bayesian Video Matting Using Motion Based Segmentation

Video matting attempts to extract foreground images from an image sequence, as well as the alpha-mattes that describe their transparency. This papers presents an automated video matting framework using motion-based segmentation and bayesian matting. We build upon existing techniques for layer segmentation based on optical flow to generate a coarse trimap, by identifying image pixels as belonging to one of background, foreground or unknown. Furthermore, we estimate occluded background for each frame by warping background pixels from neighboring frames. Bayesian methods for video matting are used for identifying detailed alpha-matte and foreground based on a gaussian mixture model distribution for foreground pixels.