Depth video camera based temporal alpha matting for natural 3D scene generation

We present a new method for fully automated video matting. This method uses depth information acquired by a depth camera to automatically compute trimaps. Trimaps segment an image into three nonoverlapping regions (foreground, background, and unknown) and generation of a highly accurate trimap is one of the most important tasks in natural alpha matting. We propose an adaptive approach to generate unknown regions according to the fuzziness of the foreground object. Fuzzy regions have wide unknown regions, whereas areas that contain sharp edges have narrow unknown areas. We further extend the standard closed-form matting method to optimize both spatial and temporal domains. Our results show the proposed method significantly reduces flickering artifacts and generates natural 3D scenes.