Computer Science Colloquium

This talk describes a unified framework for constructing photo-realistic 3D models of cities from aerial stereo views and close-range data. The talk will cover two main areas. In the first part of the talk, I will describe a new approach for rectification of stereo pairs. The proposed method is a projective rectification technique, and hence does not require camera calibration or an explicit knowledge of the epipolar geometry, although the fundamental matrix can be recovered as a byproduct. It also combines some nice features from Loop-Zhang (1999) and Hartley (1999) and exploits planar projective transformation (collineation) to streamline the problem. In particular, I will show that the use of collination reduces the problem of estimating a pair of matching homographies to that of estimating a single optimal homography. In the second part of this talk, I will propose a scheme for constructing 3D models of cities based on fusing multiple cues in the framework of Bayesian decision theory. I will show that Bayesian decision theory allows for fusing different local cues represented in the form of binary decision maps, while imposing a global consistency constraint in the form of a prior probability distribution. Time permitting, I describe some of the issues related to generation of optimal meshes and ortho-textures for photo-realistic representation. I will conclude the talk by demonstrating some simulation results and some interesting applications.

Dr. Hassan Foroosh is from University of California, Berkeley where he is a Senior Research Scientist with the Dept. of Electrical Engineering and Computer Science. He is the principal research scientist and technical coordinator of a Multi-disciplinary University Research Initiative (MURI) project, involving four other major universities.