Mean-Field Theory and Its Applications In Computer Vision1 1 Introduction • Problem formulation • Mean-field based inference method • Strategy for incorporating different costs 2 Labelling problem Assign a label to each image pixel Object segmentation Stereo Object detection 3 Problem Formulation Find a Labelling that maximize the conditional probability 4 Inference Message Passing Move-Making • T. Minka. Expectation Propagation for Approximate Bayesian Inference, UAI, 2001 • Murphy. Loopy Belief Propagation: An Empirical Study, UAI, 1999 • Jordan et.al. An Introduction to Variational Methods for Graphical Models, ML-1999 • J. Yedidia et al. Generalized Belief Propagation, NIPS, 2001 • Besag. On the Statistical Analysis of Dirty Pictures, JRSS, 1986 • Boykov et al. Fast Approximate Energy Minimization via Graph Cuts, PAMI 2001 • Komodakis et al. Fast Approximate Optimal Solutions for Single and Dynamic MRFs, CVPR, 2007 • Lempitsky et al. Fusion Moves for Markov Random Field Optimization, PAMI, 2010 Convex Relaxations Other Algorithms • Chekuri et al. Approximation Algorithms for Metric Labelling, SODA, 2001 • M. Goemans et al. Improved Approximate Algorithms for Maximum-Cut, JACM, 1995 • M. Muramatsu et al. A New SOCP Relaxation for Max-Cut, JORJ, 2003 • RaviKumar et al. QP Relaxation for Metric Labelling, ICML 2006 • K. Alahari et.al. Dynamic Hybrid Algorithms for MAP Inference, PAMI 2010 • P. Kohli et al. On Partial Optimality in Multilabel MRFs, ICML, 2008 • C. Rother et al. Optimizing Binary MRFs via Extended Roof Duality, CVPR, 2007 5 Inference Message Passing • T. Minka. Expectation Propagation for Approximate Bayesian Inference, UAI, 2001 • Murphy. Loopy Belief Propagation: An Empirical Study, UAI, 1999 • Jordan et.al. An Introduction to Variational Methods for Graphical Models, ML-99 • J. Yedidia et al. Generalized Belief Propagation, NIPS, 2001 • Variational message passing algorithm • We focus on mean-field based inference 6 Mean-field methods • Mean-fields methods (Jordan et.al., 1999) • Intractable inference with distribution • Approximate distribution from tractable family P 7 Variational Inference • Minimize the KL-divergence between Q and P 8 Variational Inference • Minimize the KL-divergence between Q and P 9 Variational Inference • Minimize the KL-divergence between Q and P 10 Variational Inference • Minimize the KL-divergence between Q and P 11 Markov Random Field (MRF) • Graph: • A simple MRF Product of potentials defined over cliques 12 Markov Random Field (MRF) • Graph: • In general Un-normalized part 13 Energy minimization • Potential and energy 14 Variational Inference Entropy of Q Expectation of cost under Q distribution 15 Naïve Mean Field • Family : assume all variables are independent 16 Variational Inference • Shannon’s entropy decomposes 17 Variational Inference • Stationary point solution • Marginal update in mean-field • Normalizing constant: 18 Variational Inference • Marginal for variable i taking label l 19 Variational Inference • Marginal for variable i taking label l • An assignment of all variables in clique c 20 Variational Inference • Marginal for variable i taking label l • An assignment of all variables in clique c • An assignment of all variables apart from x_i 21 Variational Inference • Marginal for variable i taking label l • An assignment of all variables in clique c • An assignment of all variables apart from x_i • Marginal distribution of all variables in c apart from x_i 22 Variational Inference • Marginal for variable i taking label l • An assignment of all variables in clique c • An assignment of all variables apart from x_i • Marginal distribution of all variables in c apart from x_i • Summation evaluates the expected value of cost over distribution Q given that x_i takes label l 23 Simple Illustration Naïve mean-field approximation 24 Mean-field algorithm • Iterative algorithm • Iterate till convergence • Update marginals of each variable in each iteration 25 Q distribution 26 Max posterior marginal (MPM) • MPM with approximate distribution: • MAP solution / most likely solution • Empirically achieves very high accuracy: 27 Structured Mean Field • Naïve mean field can lead to poor solution • Structured (higher order) mean-field 28 How to make a mean-field algorithm • Pick a model • Unary, pairwise, higher order cliques • Define a cost • Potts, linear truncated, robust PN • Calculate the marginal • Calculate the expectation of cost defined 29 How to make a mean-field algorithm • Use this plug-in strategy in many different models • Grid pairwise CRF • Dense pairwise CRF • Higher order model • Co-occurrence model • Latent variable model • Product label space 30