In order to segment visual objects in images, it is reasonable to use shape priors in addition to appearance characteristics like colour and texture. At the same time supervised and unsupervised learning is an essential prerequisite for applications of such models.

One possible route towards this goal is to adapt probabilistic shape priors from level set based methods and to combine them with Markov Random Fields. In this talk we are going to present such a model which allows to pose all variants of recognition tasks as well as all variants of learning tasks in a closed fashion and to solve them approximately.

A second possible route is to model shape priors in a distributed way by adapting methods of texture modelling. We will present some first ideas on this track.