Depth Recovery from Stereo Matching Using Coupled Random Fields
Year: 2010
Type: dissertation
Abstract: The depth of objects in 3-D scene can be recovered from a stereo image-pair by finding correspondences between the two views. This stereo matching task involves identifying the corresponding points in the left and the right images, which are the projections of the same scene point. The difference between the locations of the two corresponding points is the disparity, which is inversely related to the 3-D depth. In this thesis, we focus on Bayesian techniques that constrain the disparity estimates. In particular, these constraints involve explicit smoothness assumptions. However, there are further constraints that should be included, for example, the disparities should not be smoothed across object boundaries, the disparities should be consistent with geometric properties of the surface, and regions with similar colour should have similar disparities. The goal of this thesis is to incorporate such constraints using monocular cues and differential geometric information about the surface.To this end, this thesis considers two important problems associated with stereo matching; the first is localizing disparity discontinuities and second aims at recovering binocular disparities in accordance with the surface properties of the scene under consideration. We present a possible solution for each these problems. In order to deal with disparity discontinuities, we propose to cooperatively estimating disparities and object boundaries. This is motivated by the fact that the disparity discontinuities occur near object boundaries. The second one deals with recovering surface consistent disparities and surface normals by estimating the two simultaneously. 129
Source: HAL
Author: Narasimha, Ramya
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2275306325
