João Paulo Costeira "Correspondence problems in Computer Vision"

The "correspondence problem" is referred in the area of computer vision as the general process of associating points or patterns in two or more images. Obtaining the correspondences between image points in stereo pairs or in video sequences is the fundamental step to allow the 3D reconstruction of a scene from images and to being able to recognize or track objects as they move. Formulated as a combinatorial optimization problem it leads, frequently, to NP-hard problems that are very hard to tackle. We will present approaches that rely on the intrinsic properties of the 3D imaging process (low rank) and more recent methods that exploit availability of "training data" to overcome the complexity of the search.

Sponsors