Context First

We propose a probabilistic model that captures contextual 4 4 information in the form of typical spatial relationships between regions 5 5 of an image. We represent a region’s local context as a combination of the 6 6 identity of neighbouring regions as well as the geometry of the neighbour7 7 hood. We subsequently cluster all the neighbourhood configurations with 8 8 the same label at the focal region to obtain, for each label, a set of con9 9 figuration prototypes. We propose an iterative procedure based on belief 10 10 propagation to infer the labels of regions of a new image given only the 11 11 observed spatial relationships between the regions and the hitherto learnt 12 12 prototypes. We validate our approach on a dataset of hand segmented 13 13 and labelled images of buildings. Performance compares favourably with 14 14 that of a boosted, non-contextual classifier. 15 15