About My Project

I am on Professor Bajcsy's (pronounced bye-chi) research team on tele-immersion. The tele-immersive system is a set of cameras arranged around a cube, where a person stands. The cameras reconstruct a 3D image of the scene inside the cube, which can be sent to another computer over a (fast) internet connection. In other words, the goal is to develop a sort of 3D teleconferencing. For more information, see the project's website .

My project is to work on facial detection and face tracking. That is, I am working towards making the tele-immersive system detect and track faces. I'm using a package called Open CV, which has a variety of functions useful for vision research. Face detection is relatively easy--the challenge is to make robust face tracking that's fast enough for real-time video. Our goal is 25 frames/second.


We didn't make the 25 fps goal, although we were running on a slower computer.

Final Paper

I forewarn you, this paper assumes that you know some about computer vision. There's an overview of the system that should be accesiable, but we refer to various algorithms without explaining them (otherwise our paper would be very long). Now that you're warned, the paper is here:
Our final paper

Presentation Slides

I think this is a bit more accessible than the final paper. It goes less in depth, though, and parts of it were intended to be explained by the presenter. Nevertheless, it can be useful, especially paired with the paper:
Our presentation

A picture of face tracking results

Feature points with filtering. The pink points are those that were filtered away in the next frame. Note how they are generally in more homogenous regions or off the face.


Time for face detection: 350ms
Time for face tracking: 40ms
Time for feature detection: 130ms
Time for feature tracking: 20ms
Time for feature correspondence: 25ms
Time for pose estimation: negligeable