MIT Department of Electrical Engineering & Computer Science

DATE: MONDAY, APRIL 24, 1995

TIME: TALK AT 3:00
Refreshments at 2:45

PLACE: NE43-518

Representations for Visual Scene Reconstruction

Richard Szeliski
Cambridge Research Lab
Digital Equipment Corporation

The recovery of geometric information from multiple views is one of the classic problems in computer vision. Traditional applications have focused on recovering information for robotic tasks such as obstacle avoidance, grasping, and recognition. More recent applications focus on creating detailed realistic 3-D scene models for virtual reality, telepresence, and 3-D graphics applications. In this talk, I will present some recent results in these areas. and discuss representational choices underlying these techniques. To begin, I will show how to create large, high-resolutions images by panning a camera across the scene (e.g., for scanning a whiteboard or creating an outdoor panorama). Then, I will discuss some techniques for extracting scene depth from an uncalibrated moving camera, with applications to "view interpolation". Finally, I will discuss the more general 3-D scene and object recovery problem, which involves addressing some difficult issues in general-purpose 3-D shape representations.

HOST: Prof. Rodney Brooks