Volker Willert and Julian Eggert, "Belief Propagation in Spatiotemporal Graph Topologies for the Analysis of Image Sequences", Proceedings of the VISAPP, 2010.

Belief Propagation (BP) is an efficient approximate inference technique both for Markov Random Fields (MRF) and Dynamic Bayesian Networks (DBN). 2DMRFs provide a unified framework for early vision problems that are based on static image observations. 3D MRFs are suggested to cope with dynamic image data. To the contrary, DBNs are far less used for dynamic low level vision problems even though they represent sequences of state variables and hence are suitable to process image sequences with temporally changing visual information. In this paper, we propose a 3D DBN topology for dynamic visual processing with a product of potentials as transition probabilities. We derive an efficient update rule for this 3D DBN topology that unrolls loopy BP for a 2D MRF over time and compare it to update rules for conventional 3D MRF topologies. The advantages of the 3D DBN are discussed in terms of memory consumptions, costs, convergence and online applicability. To evaluate the performance of infering visual information from dynamic visual observations, we show examples for image sequence denoising that achieve MRF-like accuracy on real world data.