Ali Azarbayejani

PhD in Media Arts and Sciences

Research interests
Vision, computer vision, dynamic estimation, probabalistic modeling

Advisor: Alex Pentland

Education
Ph.D. Media Art and Sciences
Massachusetts Institute of Technology, 1997

S.B. Aerospace Engineering
Massachusetts Institute of Technology, 1988

S.M. Aerospace Engineering
Massachusetts Institute of Technology, 1991

S.M. Electrical Engineering and Computer Science
Massachusetts Institute of Technology, 1991

Research Description

My current research concerns the interaction of multiple levels and types of representation in computer vision and 3-D estmation. I am currently studying this topic in the context of tracking and interpreting human motion using probabalistic modeling and estimation.

The picture above shows my automatic blob tracking system in action --- the left is a frame from one of the stereo cameras with feature tracks and current blob features overlaid, the right is a view of the 3-D self-calibration and reconstructed trajectories of head and hands (from above) (sorry about the pictures, they were dithered from postscript). The system allows me to self-calibrate the stereo rig and recover rough location, orientation, and shape of tracked objects. My system uses two stock SGI INDYs and recovers head and hands at around 10 frames per second.

My previous research and ongoing interest centers around the modeling and feature-based estimation of motion, structure, and camera parameters from motion video. This work has been applied to vision-based head tracking, video-based model-building, and stereo camera self-calibration.

Here is an example of what Hollywood needs computer vision for. Currently production houses use human labor to twiddle 3-D parameters of camera motion, focal length, and shapes to get CG models aligned to imagery. What takes days for a person takes seconds for a computer given the image measurements, which can be done mostly automatically plus maybe a few hours time of a person.

In the above example, 20 seconds of video were analyzed with my feature-based 3-D geometry estimator allowing 3-D scene geometry and texture maps to be recovered. Two principal applications are shown on the right --- the first is extracting texture-mapped models for use elsewhere, the second is aligning 3-D models with moving imagery accounting properly for camera motion, perspective distortion, and occlusions. This is collaborative work with Bradley Horowitz, Tinsley Galyean, and Alex Pentland and a patent will issue in 1996.

Publications

Recursive Estimation of Motion, Structure, and Focal Length
Ali Azarbayejani and Alex Pentland
IEEE Trans. Pattern Analysis and Machine Intelligence, June 1995.

Visually Controlled Graphics
Ali Azarbayejani, Thad Starner, Bradley Horowitz and Alex Pentland
IEEE Trans. Pattern Analysis and Machine Intelligence, June 1993.

Model-based Vision Navigation for a Free-flying Robot
Ali J. Azarbayejani
Master's Thesis, MIT, 1991.

Recursive Estimation for CAD Model Recovery
Ali J. Azarbayejani, Tinsley Galyean, Bradley Horowitz and Alex Pentland
2nd CAD-based Vision Workshop, 1994.

Interests
Jazz, trumpet, electric bass, cycling, skiing, motorcycling

Birthplace
Royal Oak, Michigan

A favorite book
A favorite book

Most wishes to have dinner with
Most wishes to have dinner with

Personal Web Page

Vismod Home Page