|
|
MIT Electrical Engineering and Computer Science
EECS Event |
Monday, September 18, 2000
4:00 PM (refreshments 3:45)
Edgerton Hall, Room 34-101
EECS Colloquium
Abstract
Environment capture, or "geometric modeling" -- acquiring a representation of an object in a form useful for computer simulation -- is an essential first step in visualization, simulation, and computer-aided design. Researchers have developed a variety of techniques for extracting the necessary information from photographs. Historically, proposed techniques have been either algorithmic (limited to small-scale, restricted settings), or interactive (limited by the human operator's skill and energy). To date, no single system has achieved both automation and scaleable, end-to-end operation.
We break the logjam by using the scale of the problem to advantage. We assume that the urban scene to be captured exhibits regular structure in the form of sets of parallel lines and identifiable window and building corners. We acquire thousands of images of the scene, and register and combine them using algorithms which incorporate geometric duality, geometric statistics, graph propagation, consensus techniques and numerical optimization to extract high-fidelity geometric models.
The scale and generality of the problem impose significant constraints on any proposed system design. We discuss these constraints and their implications, then present our model capture system. Eliminating the human is a significant challenge from both engineering and research standpoints, and the effort has led to some powerful new techniques. In contrast to the prevailing view that human intervention always improves quality, we give examples of situations in which our automated system outperforms a human operator. We describe the status of the project and show some recent results from the MIT campus.