Many of the most interesting problems in computer vision are difficult because they are inherently ill-posed and the corresponding algorithms have to cope with outliers, as well as missing or incomplete data. One of the ways to tackle such problems is to impose a set of constraints encoding a prior knowledge about feasible solutions. A popular method to build such constraints is to utilise an observation model formed either analytically through the analysis of the observation process or empirically utilising representative samples from population of observations.

The objective of this talk will be to describe speaker current work on model based visual information processing. Although examples will be drawn from a number of fundamental computer vision problems, including segmentation, registration and tracking showing their diverse practical applications, the main emphasis will be put on facial expressions and human body posture analysis. In case of the facial expression analysis a whole processing pipeline will be explained including feature extraction, model building, and classification as well as validation based on a purpose build facial expression database.