Learning Low-Complexity Face Models for Tracking and Recognition

The human face is an important topic of research in computer vision as people use faces to communicate in various ways. We are interested in developing 3-D mesh-based face models which can be used to track and recognize faces. This thesis will contribute techniques for the computation and application of low-complexity face geometry models with functional subspaces.

The following is a summary of the major contributions of this dissertation:

  1. A better method to automatically learn a simple 3-D face model from training stereo image sequences.
  2. A novel algorithm to simultaneously compute the expression and identity functional subspaces of the face from training data.
  3. A contribution to tracking the 3-D pose, position, and facial expression in a monocular image sequence using our 3-D model.
  4. The use of the low-dimensional identity subspace of our model to recognize faces.
  5. The ability to recognize facial expressions using the action subspace of our model.