Object tracking is an important tool for rotoscoping. Tracking using snakes is one obvious approach to the rotoscoping problem: The snake tracks the object outline and then determines the alpha-channel values. However, without detailed object-shape and -motion models, current snake formulations do not reliably track occlusions in natural scenes.
We have developed a new external-energy term for snakes which allows occlusion tracking without detailed object models. Our paper shows how our image model improves tracking in cluttered scenes. Finally, our paper describes how a high-quality rotoscoping mask is created using our image model both to drive the snake and to refine the silhouette of the snake track to match the object.
For a copy of our paper, please see either the PDF version (440k) or the Postscript version (14M) .
Tracking BoundariesOur new external-energy term for snakes, called radial cumulative similarity (RCS), captures foreground attributes, as does patch matching; captures the occlusion structure, as does edge matching; and ignores background attributes, as does robust matching.
|
Automatic Rotoscoping
RCS snakes, along with their associated RCS profiles, can be used for semi-automatic rotoscoping of foreground objects. The artist draws a closed-contour snake in the first frame of the sequence. The full length of the snake should lie completely inside the object's occluding contour. This closed-contour snake is propagated through the entire sequence using RCS snake tracking. Once this is done, the selected object is automatically rotoscoped out of the sequence. This is done using a continuous-valued alpha-channel sequence created from the snake position and from the RCS profiles along the snake. The separate pieces from automatic rotoscoping the inch worm and the lip sequences are available: the inch worm and its surroundings are separated, as are the lips and the inside of the mouth. |