Figure 1: (left) Screenshots from happy testing clips, for both RBFN and ANN retargeting systems, (right), Screenshots from angry testing clips. The comparison with the reference video footage shows that both generated similar output, although with more data we could expect the ANN to perform better.
AbstractMapping the motion of an actor's face to a virtual model is a difficult but important problem, especially as fully animated characters are becoming more common in games and movies. Many methods have been proposed but most require the source and target to be structurally similar. Optical motion capture markers and blendshape weights are an example of topologically incongruous source and target examples that do not have a simple mapping between one another. In this paper, we created a system capable of determining this mapping through supervised learning of a small training dataset. Radial Basis Function Networks (RBFNs) have been used for retargeting markers to blendshape weights before but to our knowledge Multi-Layer Perceptron Artificial Neural Networks (referred to as ANNs) have not been employed in this way. We hypothesized that ANNs would result in a superior retargeting solution compared to the RBFN, due to their theoretically greater representational power. We implemented a retargeting system using ANNs and RBFNs for comparison. Our results found that both systems produced similar results (figure 1) and in some cases the ANN proved to be more expressive although the ANN was more difficult to work with.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.