A Tiered Move-making Algorithm for General Pairwise MRF’s
Vibhav Vineet, Jonathan Warrell and Philip Torr
In: CVPR 2012, 18-20 June 2012., Rhode Island.
A large number of problems in computer vision can bemodeled as energy minimization problems in a markov ran-dom field (MRF) framework. Many methods have been de-veloped over the years for efficient inference, especially inpairwise MRFs. In general there is a trade-off between thecomplexity/efficiency of the algorithm and its convergenceproperties, with certain problems requiring more complexinference to handle general pairwise potentials. Graph-cuts based α-expansion performs well on certain classes ofenergies, and sequential tree reweighted message passing(TRWS) and loopy belief propagation (LBP) can be used fornon-submodular cases. These methods though suffer frompoor convergence and often oscillate between solutions.In this paper, we propose a tiered move making algo-rithm which is an iterative method. Each move to the nextconfiguration is based on the current labeling and an op-timal tiered move, where each tiered move requires oneapplication of the dynamic programming based tiered la-beling method introduced in Felzenszwalb et. al. .The algorithm converges to a local minimum for any gen-eral pairwise potential, and we give a theoretical analy-sis of the properties of the algorithm, characterizing thesituations in which we can expect good performance. Weevaluate the algorithm on many benchmark labeling prob-lems such as stereo, image segmentation, image stitchingand image denoising, as well as random energy minimiza-tion. Our method consistently gets better energy values thanα-expansion, LBP, quadratic pseudo-boolean optimization(QPBO), and is competitive with TRWS.
|EPrint Type:||Conference or Workshop Item (Paper)|
|Project Keyword:||Project Keyword UNSPECIFIED|
|Deposited By:||Sunando Sengupta|
|Deposited On:||16 June 2012|