Natural Matting of Complex Scenes

Image matting has been used for a long time in the movie industry to combine several picture elements into a final image. Typically an actor or similar is considered the foreground and extracted from the filming environment. The resulting matte is then used to place the foreground onto a different background by applying alpha blending. One of the most common techniques for the matting procedure is the use of blue screens, this gives the background a constant color that is easy to locate and separate from the foreground. The use of blue screens is however limited to controlled studio environments which is not always practical due to e.g. physic or economic constraints. To overcome these problems natural image matting has to be used, i.e. where images have natural and non-constant backgrounds. Another use for natural image matting is to insert objects into a filmed scene behind a moving foreground as demonstrated in figure 1 where a wall painting and a flower have been inserted behind a moving actor. Several approaches have been proposed to accomplish natural image matting, but difficulties arises in complicated transitions between foreground and background. Complicated transitions could for example be the hair of an actor, grass, or other non-trivial objects. Chuang et al. [2] have proposed a Bayesian framework to accomplish stable matting of such complex scenes in images. This technique was later extended by Chuang et al. [1] using optical flow and background estimation to make matting of video sequences practical. This paper will describe the techniques used by Chuang et al. [2, 1] and show the possibilities when using natural image matting in the movie industry.