Algorithms for Markov Random Fields in Computer Vision.pdfVIP

Algorithms for Markov Random Fields in Computer Vision Dan Huttenlocher November, 2003 Joint work with Pedro Felzenszwalb Random Field Broadly applicable stochastic model – Collection of n sites S – Hidden variable xi at each site i 1 2 – Label set L ? Each site takes on label l ∈L 3 4 – Neighborhood system N ? Ni neighbors of site i ? Explicit dependencies between neighbors Graphical model with undirected edges – Graph G S,N – Ni set of nodes with edges incident on i 2 Markov Random Field MRF Random field with Markov property P x | x P x | x i S\i i Ni – Where S\i denotes set S excluding element i Standard simplification abuse of notation – Probability of r.v. x taking on value v, P x v i i abbreviated as P xi Conditional probabilities depend only on neighborhood – Probability of x taking on some value same i given all other nodes as given just neighbors 3 MRF’s for Low Level Vision Grid graph – Sites are pixels; up, down, left, right neighbors – Neighborhood enforces spatial coherence – Observed value y at each site pixel i Applies to many pixel-oriented problems – Naturally expressed as posterior probability of labels given observations, P x|y ? Stereopsis, labels are depths disparities ? Optical flow, labels are motion vectors ? Restoration, labels are intensities colors – [Geman Geman, 1984] 4 MRF Stereo Given two images, estimate depth at each pixel y3 y4 y1 y2 x x 3 4 x x 1 2 Depths Tens of labels 5 MRF Motion Given two images, estimate motion vector at each pixel y3 y4 y1 y2 x x 3 4 x x 1 2 Flow vectors Hundreds of labels 6 MRF Image Restoration Given image corrupted by noise, estimate original image – Intensity/color for each pixel y3 y4 y1 y2 x x 3 4 x x 1 2 Intensities Hundreds of labels 7 MRF’s for High Level Vision Given image, estimate location of object – Pictorial structure model ? Parts represented as local image patches ? Spring-like connections between pairs of parts – Non-MRF formulation [FischlerElschlager, 1973] x 10 x 1 x