Linear的 Time Methods for Propagating Beliefs.pptVIP

Linear Time Methods forPropagating BeliefsMin Convolution, Distance Transforms and Box Sums Daniel HuttenlocherComputer Science DepartmentDecember, 2004 Problem Formulation Find good assignment of labels xi to sites i Set L of k labels Set S of n sites Neighborhood system N?S?S between sites Undirected graphical model Graph G=(S,N) Hidden Markov Model (HMM), chain Markov Random Field (MRF), arbitrary graph Consider first order models Maximal cliques in G of size 2 Problems We Consider Labels x=(x1,…,xn), observations (o1,…,on) Posterior distribution P(x|o) factorsP(x|o) ? ?i?S?i(xi) ?(i,j)?N?ij(xi,xj) Sum over labelings?x(?i?S?i(xi) ?(i,j)?N?ij(xi,xj)) Min cost labelingminx(?i?S?’i(xi)+?(i,j)?N?’ij(xi,xj)) Where ?’i = -ln(?i) and ?’ij = -ln(?ij) Computational Limitation Not feasible to directly compute clique potentials when large label set Computation of ?ij(xi,xj) requires O(k2) time Issue both for exact HMM methods and heuristic MRF methods Restricts applicability of combinatorial optimization techniques Use variational or other approaches However, often can do better Problems where pairwise potential based on differences between labels ?ij(xi,xj)=?ij(xi-xj) Applications Pairwise potentials based on difference between labels Low-level computer vision problems such as stereo, and image restoration Labels are disparities or true intensities Event sequences such as web downloads Labels are time varying probabilities Fast Algorithms Summing posterior (sum product) Express as a convolution O(klogk) algorithm using the FFT Better linear-time approximation algorithms for Gaussian models Minimizing negative log probability cost function (corresponds to max product) Express as a min convolution Linear time algorithms for common models using distance transforms and lower envelopes Message Passing Formulation For concreteness consider local message update algorithms Techniques apply equally well to recurrence formulations (e.g., Viterbi) Iterative local update sch