Face view synthesis using a single image

Face view synthesis involves using one view of a face to artificially render another view. It is an interesting problem in computer vision and computer graphics, and can be applied in the entertainment industry for animated movies and video games. The fact that the input is only a single image, makes the problem very difficult. Previous approaches learn a linear model on pair of poses from 2D training data and then predict the unknown pose in the test example. Such 2D approaches are much more practical than approaches requiring 3D data and more computationally efficient. However they perform inadequately when dealing with large angles between poses. In this thesis, we seek to improve performance through better choices in probabilistic modeling. As a first step, we have implemented a statistical model combining distance in feature space (DIPS) and distance from feature space (DFFS) for such pair of poses. Such a representation leads to better performance. As a second step, we model the relationship between the poses using a Bayesian network. This representation takes advantage of the sparse statistical structure of faces. In particular, we have observed that a given pixel is often statistically correlated with only a small number of other pixel variables. The Bayesian network provides a concise representation for this behavior reducing the susceptibility to over-fitting. Compared with the linear method, the Bayesian network more accurately predicts small and localized features.