This thesis treats two topics in model-based coding; coding of facial textures (Part 1) and real-time head tracking (Part 2).

In Part 1, it is shown that a face image can be efficiently coded using a parameterization based on geometrical normalization followed by Karhunen-Loeve transformation (KLT). The resulting parameterization is shown to be convex, and can be used for coding: It improves both the measured and perceived quality of face images compared to only using eye matching normalization followed by KLT. A block based version of the coder improves the performance dramatically both in terms of quality and complexity. By distributing the bits unevenly over the face, quality can be further improved. Comparisons with JPEG show an improvement of 8 dB.

In Part 2, a real-time head tracker that is robust to large rotations is described. The system uses a large number of automatically selected feature points, constrained by dynamically estimated structure. The structure from motion (SfM) algorithm described by Azarbayejani and Pentland in 1995 is used. The algorithm is first examined for planar objects, and then extended to manage points not visible in the first frame. The extended SfM method is the basis for the head tracker: A texture mapped three-dimensional head model is created, and 24 feature points on the surface of this model are automatically selected and tracked. The trajectories of the tracked feature points are forwarded to the SfM algorithm, which in turn provides estimates of the location of the feature points in the next frame. By adaptively updating the texture, the extended SfM algorithm can be used to track points on, e.g., the side of the head, which improves the range of the tracker. A reinitialization procedure that uses data from the original initialization is also presented. The complete system runs at full frame rate (25/30 Hz) and is evaluated on both real and synthetic data.