Markov chain Monte Carlo for integrated face image analysis

This PhD thesis is about the integration of different methods to fit
a statistical model of human faces to a single image. I propose to take
a probabilistic view on the problem and implement and evaluate an integrative
framework for face image explanation based on a class of methods known as
Data-Driven Markov Chain Monte Carlo. The framework is based on the
propose-and-verify architecture of the Metropolis-Hastings algorithm.
Probabilistic inference replaces traditional optimization methods and
conceptually shifts the goal of face explanation from obtaining the optimal
parameter set to extracting measures of the posterior distribution. The
probabilistic view opened the process for deeper insights like the need of
a background model and richer likelihood models. Within this framework,
different methods are implemented and evaluated specifically for face image
explanation with the 3D Morphable Model and face and feature point detection.
The Markov Chain Monte Carlo integration method is able to algorithmically
reproduce existing fitting algorithms as well as capable of dealing with
unreliable and differently shaped information sources. The integration of
Bottom-Up information into the adaption process leads to more robust results
than a simple feed-forward combination of the methods and culminates into
a fully automatic face image explanation method, independent of user-provided
initialization. A full-system application leads to a fully automatic and
general face recognition application with state of the art results.