Author:

Keywords:

PSI_VISICS

Abstract:

This doctoral thesis concerns the modeling of the human face using generic models and in particular, 3-D morphable models (3DMMs). A 3DMM is a continuous parameterization of the 3-D shape and texture variations within an object class. It establishes a mapping between a low-dimensional parameter space, and a high-dimensional space of textured 3-D models. In addition to this parameterization, 3DMMs contain statistical information about the object class in the form of a probability density function (PDF) defined on the parameter space. We begin by describing how such a model can be constructed from a set of 3-D scans of human faces. We show how the model can be used to reconstruct a person’s face from either 3-D data or images, even when the input data is partially occluded. A fully automatic system is described for fitting a facial 3DMM to a single image, without manual annotations. Since the accuracy of the output is limited by the expressiveness of the model, a substantial amount of work has gone into designing methods for generalizing the model outside the training set. We explore different types of part-based models and design an algorithm for automatically finding the optimal facial parts for reconstructing novel faces. It turns out that the optimal parts differ substantially from what we might expect intuitively. Finally, we explore some of the many applications of 3DMMs, such as automatic face recognition and lip synchronization. We also illustrate how such models can be used in the field of psychology.