Dr. Jean Ponce
Ecole Normale Superieure, Paris, France.
I will present in this talk some recent work aimed at understanding the geometry of very general imaging systems that may include refractive and reflective interfaces. A key observation is that, irrespective of the optics, any image is formed by intersecting some sensor array with a 2D family of straight lines (the impinging lightrays) forming a so-called line congruence. Line geometry is thus the key to a unified theoretical framework for 3D computer vision. Concretely, I will revisit multi-view geometry using the concurrent lines variety formed by n-tuples of lines that meet at a point, leading to new characterizations of geometric consistency for point correspondences. I will also use geometric and analytical models of line congruences to extend this approach to completely general camera models, including pinhole, two-slit, and catadioptric cameras in a unified way. In particular, I will present novel structure-from-motion and self-calibration algorithms for two-slit cameras, and discuss completely new camera designs based on the proposed framework.
Hosted by: Center for Research in Computer Science
Read More