Abstract

Traditionally, the matching problem in stereo vision has been formulated as an ill-posed problem. However, it has been shown that the matching problem can be well-posed in scenes with no occlusions or depth discontinuities. Unfortunately most real scenes do not obey these constraints. We overcome this by finding regions within the images in which the matching problem is well-posed. That is, we find image regions in which there are no occlusions or depth discontinuities. In general, a unique set of such regions does not exist. However, we will demonstrate that in almost all cases these regions can be found efficiently. Therefore the matching problem can be well-posed in almost all cases. In order to find these corresponding regions we transform the problem from finding 2D image regions into identifying 3D surfaces. We have developed a method of 3D surface extraction which uniquely identifies correct 3D surfaces from a set of potential surfaces. In order to test the method we have built ...

Citations