Wide baseline matching with applications to visual servoing

Abstract

Matching is a fundamental problem in computer vision. In one formulation of the matching problem, the task is to determine what point in one image that corresponds to a given point in another image of the same scene. If sufficiently many correspondences between the images are known, the 3D structure of the scene may be recovered. A system capable of performing this task would have applications in robotics, computer vision for the entertainment industry and media, and 3D modeling from multiple views. Stereo matching is a restricted version of the image matching problem, in which the viewpoints of the different images are required to have a very small separation. In such cases, corresponding points will have almost the same position in both images, and the complexity of the matching problem is reduced. Furthermore, a potential correspondence between two images may be verified by requiring that the image neighbourhoods of the points are similar to one and another. As the viewpoint difference between the images is assumed to be small, this verification can be made using cross-correlation of image data. Many