Reusing Traditional Animation

Project Description

Traditional (cartoon) animation is notable for its expressiveness and style. It is difficult to produce, however. Since a large body of traditional animation already exists, it may be possible to incorporate this body into a library that can be used to generate new animation having the same qualities that make the medium so appealing. This work presents components of a system to do exactly that.

Also, sometimes the realism of motion capture or dynamic simulation is not wanted. One can imagine learning environments, particularly for children, in which traditionally animated characters with caricatured or non-human features serve as a better representation for an agent. Authoring, editing, and re-using traditional animation are some of the most difficult and time-consuming animation tasks. We have developed ways of generating novel traditional animation from existing traditional animation. Our method is model-free, i.e., no a priori knowledge of the drawing or character is required.

There are three primary challenges in building the system. The first is to put the characters into a form in which the character is nicely separated from the background. Much older cartoon data suffers from noise due to changes in lighting as the cel animations were transferred to film, contamination of the cel from one use to another as it was filmed, and degradation of the animation before being transferred to an archival format. These factors make the segmentation problem difficult. We solve the segmentation problem using support vector machines. The second challenge is then to reuse the animation. We address that problem by providing semi-automatic methods to generate an inbetween of two key images. We solve this problem using non-rigid elastic deformation of the images and radial basis functions. The final challenge lies in resequencing the library of cartoon data. We treat this lbirary as a low-dimensional manifold embedded in high-dimensional space. By recovering the manifold structure of the set, we can generate novel animation. Additional processing is needed to improve its appearance, however, and by treating the problem as a scattered data interpolation problem we have successfully generated novel sequences.