| dc.description.abstract | Games and movies use fully animated characters more and more. Animating these characters manually is a lot of work, and subtle expressions can be missed. Accurate automated mapping of the actor’s face to a virtual character is essential to have. To be able to do this, the mapping between an actor’s face and the corresponding virtual character needs to be calculated. In a previous paper, an artificial neural network was used, which led to the suggestion to try deep neural networks. No paper has been published using deep neural networks for this problem yet. The usage of a convolutional neural network (CNN), deep belief network (DBN), and recurrent neural network (RNN) is discussed to conclude that RNN shows the most promise of the three. The reason behind this is that RNN can use the prior and future frames to predict the current, which in theory would be helpful for this problem. A radial basis function network or RBFN with Hardy multi-quadric kernel is used as a comparison against RNN. Figure 1 shows a comparison for all characters used for a frame in the angry test set. In that figure, RBFN has the eyes and mouth more closed than RNN has. To determine how RNN compares to RBF a survey was held, and a cost function was made for machine learning results. The survey results show that RBFN is significantly better for three out of the five expressions, all the characters, and in general. The machine learning results show that RBFN performs better as well, with the most significant difference in the mouth area. However, only part of the possibilities of RNN is explored. Various options to improve the results of RNN are listed in the future work section. It is likely that with future research RNN can be improved to be equal or better than RBFN. | |
