ปรียกร ทิพยวัย. A Coarse-and-Fine Bayesian Belief Propagation for Correspondence Problems in Computer Vision. Doctoral Degree(Electrical and Computer Engineering ). King Mongkut's University of Technology Thonburi. : King Mongkut's University of Technology Thonburi, 2006.
A Coarse-and-Fine Bayesian Belief Propagation for Correspondence Problems in Computer Vision
Abstract:
Computer vision problems often involve the recovery of lost information from image projection. For example, a problem in computer vision may be to recover 3-dimensional surface information from one or more 2-dimensional images. Such problems are ill- posed and must rely on extra constraints afforded from the assumptions about the world such as surface smoothness to arrive at a solution. Optimization is required to find the best solution, given the data and constraints. Bayesian belief propagation (BBP) is an optimization method that has recently gained much interest in computer vision owing to its property of quick convergence to near optimal solutions. It works by passing messages around within a predefined neighborhood. Working iteratively, the message can pass along from one neighborhood to another; therefore generating a large neighborhood. We present the use of a multi-resolution, coarse-and-fine, pyramid image architecture to solve correspondence problem in various computer vision modules including shape retrieval through contour matching, stereovision, and motion estimation. The algorithm works with a grid matching and an inter-grid correspondence model by message passing in a Bayesian belief propagation (BBP) network. The local smoothness and other constraints are expressed within each resolution scale grid and also between grids in a single paradigm. Top-down and bottom-up matching are concurrently p&formed for each pair of adjacent levels of the image pyramid level in order to find the best matched features at each level simultaneously. The coarse-and- fine algorithm uses matching results in each layer to constrain the process in its adjacent 2 layers by measuring the consistency between corresponding points among adjacent layers so that good matches at different resolution scales constrain one another. The coarse-and-fine method helps avoid the local minimum problem by bringing features closer at the coarse level and yet providing a complete solution at the finer level. The method is used to constrain the solution with examples in shape retrieval, stereovision, and motion estimation to demonstrate its desirable properties such as rapid convergence, the abilities to obtain near optimal solution while avoiding local minima, and immunity to error propagation found in the coarse-to-fine approach.