CRF Based Region Classification Using Spatial Prototypes

This paper proposes a probabilistic model using conditional random field (CRF) for region labelling that encodes and exploits the spatial context of a region. Potential functions for a region depend on a combination of the labels of neighbouring regions as well as their relative location, and a set of typical neighbourhood configurations or prototypes. These are obtained by clustering neighbourhood configurations obtained from a set of annotated images. Inference is achieved by minimising the cost function defined over the CRF model using standard Markov Chain Monte Carlo (MCMC) technique. We validate our approach on a dataset of hand segmented and labelled images of buildings and show that the model outperforms similar such models that utilise either only contextual information or only non-contextual measures.