Charles R. Dyer

Professor
Department of Computer Sciences
University of Wisconsin-Madison
1210 West Dayton Street
Madison, Wisconsin 53706-1685 USA

E-mail: dyer@cs.wisc.edu
Telephone: (608) 262-1965
Fax: (608) 262-9777
Finger info

Ph.D., University of Maryland, 1979 (Curriculum Vitae)
Areas of Interest: Computer vision, three-dimensional shape representation, appearance modeling, view synthesis, active vision, visualization

Group Leader, Wisconsin Computer Vision Group

Program Co-Chair, CVPR '96


Research Interests

View Synthesis

The goal of this work is to develop basic tools for controlling in real-time, either autonomously or interactively, a virtual camera of a real environment. The input is a set of images or video streams, acquired from fixed or mobile cameras around a site, and the output is a panoramic visualization of the scene in which a virtual, user-controlled camera can be moved through the environment. With this technology a user could interactively navigate through a real environment, controlling a customized path of views of the site that are not predetermined by the input images. The main research question is how to adaptively combine a set of basis images to synthesize new views of the scene without 3D models or 3D scene reconstruction as an intermediate step. Recently we have developed an innovative technique, which we call view morphing, that takes two basis images and interpolates a continuous range of in-between images corresponding to views on the linear path connecting the two camera centers.

Visual Exploration

Computer vision researchers have recently started to investigate how to actively control the image acquisition process by controlling camera parameters. We have been studying how to purposefully control the position and orientation of a camera in order to dynamically adjust viewpoint based on the appearance of a three-dimensional scene. The use of real-time viewpoint-control behaviors is important for solving tasks such as exploring an unknown object in order to find specific surface markings, building a global model of an unknown shape, or recognizing an object.

By coordinating simple observer behaviors that change the appearance of surfaces in a well-defined way, we simplify the image computations required, make precise the global progress of an algorithm, and eliminate the need for accurate differential measurements of the camera motion. We believe that this active approach of moving towards viewpoints that are closely related to the geometry of the viewed objects is a very important and general one. We have used this approach to develop provably-correct algorithms for (1) moving to a side view of a surface of revolution in order to recover its shape, and (2) reconstructing the global surface of an unknown smooth, arbitrarily-shaped object.

Visualization

In the area of visualization we have developed mapping techniques capable of generating displays of all possible data objects defined in a user's algorithm, without the need for user-defined, type-specific graphics display procedures. This capability for displaying arbitrary combinations of an algorithm's data objects in a common frame of reference, coupled with interactive control of algorithm execution, provides a powerful way to understand algorithm behavior, especially interactive visual experiments with scientific data analysis algorithms. We have implemented a system called VIS-AD for experimenting with these techniques and have used it for visualizing intermediate and final results of data analysis algorithms for problems such as discriminating clouds in satellite images.


Recent Publications

Recent publications (including abstracts) by Wisconsin Computer Vision Group


Courses Taught


Current Ph.D. Students


Ph.D. Graduates


Links of Interest


My Bookmarks