Learning Markovian Dependencies from Annotated Images

In this paper we propose to model structural knowledge about scenes by a Markov random field whose conditional probabilities are learned from the spatial and topological relationships observed between regions in a set of training images. A locally consistent labelling of new scenes is achieved by relaxing the Markov random field directly, using conditional probabilities rather than a Gibbs formulation. We validate our approach on several hundreds of hand-segmented photographs of buildings.