Labeling irregular graphs with belief propagation

This paper proposes a statistical approach to labeling images using a more natural graphical structure than the pixel grid (or some uniform derivation of it such as square patches of pixels). Typically, low-level vision estimations based on graphical models work on the regular pixel lattice (with a known clique structure and neighborhood). We move away from this regular lattice to more meaningful statistics on which the graphical model, specifically the Markov network is defined. We create the irregular graph based on superpixels, which results in significantly fewer nodes and more natural neighborhood relationships between the nodes of the graph. Superpixels are a local, coherent grouping of pixels which preserves most of the structure necessary for segmentation. Their use reduces the complexity of the inferences made from the graphs with little or no loss of accuracy. Belief propagation (BP) is then used to efficiently find a local maximum of the posterior probability for this Markov network. We apply this statistical inference to finding (labeling) documents in a cluttered room (under moderately different lighting conditions).