Accelerated Sparse Feature Correspondence Resolution Using Loopy Belief Propagation with Mrf Clique Based Structure Preservation

In this paper we extend the work of (Louw and Nicolls, 2007) which proposed a novel Markov Random Field formulation for resolving sparse features correspondences in image pairs. The MRF terms can include cliques of variable sizes, and the energies are minimized using Loopy Belief Propagation. In this paper, an improved MRF topology is developed which uses a variant of the previously developed KN-means algorithm (where each mean has a specified number of neighbours). The message passing schedule is an accelerated one which converges faster than the usual parallel message update schedule, and surprisingly, often gives better correspondence results. The method is compared to other state of the art sparse feature correspondence algorithms and shown to compare well. Outliers are handled naturally within this paradigm.