Approximate MRF Inference Using Bounded Treewidth Subgraphs
Graph cut algorithms , commonly used in computer vision, solve a first-order MRF over binary variables. The state of the art for this NP-hard problem is QPBO [1,2], which finds the values for a subset of the variables in the global minimum. While QPBO is very effective overall there are still many difficult problems where it can only label a small subset of the variables. We propose a new approach that, instead of optimizing the original graphical model, instead optimizes a tractable sub-model, defined as an energy function that uses a subset of the pairwise interactions of the original, but for which exact inference can be done efficiently. Our Bounded Treewidth Subgraph (k-BTS) algorithm greedily computes a large weight treewidth-k subgraph of the signed graph, then solves the energy minimization problem for this subgraph by dynamic programming. The edges omitted by our greedy method provide a per-instance lower bound. We demonstrate promising experimental results for binary deconvolution, a challenging problem used to benchmark QPBO : our algorithm performs an order of magnitude better than QPBO or its common variants , both in terms of energy and accuracy, and the visual quality of our output is strikingly better as well. We also obtain a significant improvement in energy and accuracy on a stereo benchmark with 2nd order priors , although the improvement in visual quality is more modest. Our method’s running time is comparable to QPBO.
KeywordsGreedy Algorithm Original Graph Tree Decomposition Signed Graph Exact Inference
Unable to display preview. Download preview PDF.
- 1.Boros, E., Hammer, P.L.: Pseudo-boolean optimization. Discrete Applied Mathematics 123 (2002)Google Scholar
- 4.Rother, C., Kolmogorov, V., Lempitsky, V., Szummer, M.: Optimizing binary MRFs via extended roof duality. In: CVPR (2007)Google Scholar
- 8.Boros, E., Hammer, P.: A max-flow approach to improved roof-duality in quadratic 0 − 1 minimization. Technical report, RUTCOR (1989)Google Scholar
- 10.Hammer, P., Rudeanu, S.: Boolean Methods in Operations Research and Related Areas. Springer (1968)Google Scholar
- 13.Feige, U., Goemans, M.: Approximating the value of two power proof systems, with applications to max 2sat and max dicut. In: 3rd Israel Symposium on the Theory of Computing Systems, p. 182 (1995)Google Scholar
- 18.Garey, M., Johnson, D.: Computers and Intractability. W. H. Freeman and Company (1979)Google Scholar
- 19.Wainwright, M.J., Jordan, M.I.: Graphical Models, Exponential Families, and Variational Inference. Now Publishers Inc., Hanover (2008)Google Scholar
- 20.Lipton, R., Tarjan, R.: Applications of a planar separator theorem. SIAM Journal on Computing, 615–627 (1980)Google Scholar
- 21.Karger, D.R., Srebro, N.: Learning markov networks: maximum bounded tree-width graphs. In: SODA, pp. 392–401 (2001)Google Scholar
- 22.Shahaf, D., Chechetka, A., Guestrin, C.: Learning thin junction trees via graph cuts. In: Artificial Intelligence and Statistics, AISTATS (2009)Google Scholar
- 23.Mooij, J.M.: libDAI: A free & open source C++ library for Discrete Approximate Inference in graphical models. Journal of Machine Learning Research 11 (August 2010)Google Scholar
- 24.Raj, A., Zabih, R.: A graph cut algorithm for generalized image deconvolution. In: International Conference on Computer Vision, ICCV (2005)Google Scholar