Unconstrained face recognition using MRF priors and manifold traversing

In this paper, we explore new methods to improve the modeling of facial images under different types of variations like pose, ambient illumination and facial expression. We investigate the intuitive assumption that the parameters for the distribution of facial images change smoothly with respect to variations in the face pose angle. A Markov Random Field is defined to model a smooth prior over the parameter space and the maximum a posteriori solution is computed. We also propose extensions to the view-based face recognition method by learning how to traverse between different subspaces so we can synthesize facial images with different characteristics for the same person. This allow us to enroll a new user with a single 2D image.