April 2002
Tuesday, April 30, 2002
Seminar at 4pm, B&H (Engineering) bldg., Room 190
Organized by
the Division of Engineering
and the SHAPE
Lab.
Computer Vision Laboratory
Center for Automation Research
(CfAR)
University of Maryland, MD, USA
*Joint work with Cornelia Fermuller
As we move through our environment we continuously acquire images from different viewpoints. From the sequences of images (or videos) we are able to derive three-dimensional (3D) descriptions of the (possibly changing) scene in view. How is this achieved? The question is fundamental to the understanding of intelligence and to automation, since intelligent systems need to derive descriptions of their environment in order to interact with it. I refer to these descriptions as visual space-time.
In this talk I will describe the basics of our approach to this complex and heavily studied problem -- known under a variety of names, such as "model building", "structure from motion", "multi-view geometry"; it entails a number of processes which start with images and produce descriptions of 3D shape and movement. In the course of the exposition I will uncover inherent limitations in the computational processes involved and propose a theory for computing 3D descriptions which takes these limitations into account.
The computations start by relating the images, specifically deriving features in the images that correspond to the same 3D world entity. (a) I will show that because of statistical reasons there is uncertainty in the estimation of features. I will introduce a new uncertainty principle in visual processing. This principle predicts, for the first time, all geometric optical illusions. (b) Then I will show that the geometry of the imaging surface (the shape of the eye and distribution of photoreceptors) is related to the accuracy with which 3D shape and movement can be recovered. Using these theoretical findings and taking inspiration from nature I will outline a hierarchy of eyes and propose new camera designs for which the 3D recovery problem becomes easy. (c) If time permits, I will describe the details of algorithms motivated by Grenander's pattern theory and show the beginnings of a new technical discipline, termed Harmonic Computational Geometry, that allows the implementation of feedback processes in the 3D recovery problem.
I will conclude with applications to 3D video, camera networks, video manipulation and 3D video editing, and new interfaces and I will show many experimental results with real world imagery.
Last Updated: April 26, 2002