Face Modelling and Recognition Tutorial ( part II )
Thomas Vetter and Sami Romdhani

Adobe Portable Document Format (pdf) [11M]
Zip file containing the PDF presentation and the movies (zip) [40M]

This tutorial is the second part of the Face Recognition Tutorial presented at the 8th European Conference on Computer Vision, on the May 10, 2004, by Thomas Vetter and Sami Romdhani.
The first part of the tutorial was presented by Wen-Yi Zhao.

In this part of the tutorial, we explain a method for modelling human face images at any pose and under any illumination. This method is called the 3D Morphable Model. In the first part of the tutorial, we motivate the representation used by the 3D Morphable Model, we describe some of its features and explain its construction. In the second part, we detail how to achieve accurate pose and illumination invariant face recognition by use of the 3D Morphable Model. A centre piece of the recognition system is the Fitting algorithm used to register a face image with the model, thereby extracting the identity parameters and the imaging parameters explaining the input image. We review and compare five fitting algorithms:

  • Stochastic Newton Optimisation,
  • Active Appearance Model Fitting,
  • Inverse Compositional Image Alignment,
  • 3D Morphable Model Inverse Compositional Image Alignment, and
  • Linear Shape and Texture 3D Morphable Model Fitting.
The major difference between these algorithms is their speed and their accuracy. It turns out that there is a trade-off between speed and accuracy: the fastest algorithm (several frames per second) is one of the least accurate and the slowest algorithm (4.5 min per image) is the most accurate.


Sami Romdhani