| Research Projects |
|
| Weakly labeled data |
|
|
Some thoughts on associative learning with
perceptual state discovery and an EM-based algorithm are in this
paper. This paper is more less rehashing the ideas from Trial By Eire, but it is more accurate and
is properly grounded. |
|
| Trial By Eire |
|
|
My involvement with this was the utterance
learning algorithm. There are two - simpler and faster which is actually
used and another
one, which is based on EM |
|
| K9.0 |
|
|
A short
talk about models of animal conditioning at our behaviors
group meeting which I gave just to sort things out for myself, so, it is pretty much
novelty-free. |
|
| Conversation modeling |
|
|
Tony
and I have been working on a system for topic detection in a casual
conversation. Here
is a wordy write-up on that. |
|
| SCFG-based Surveillance System |
|
|
In the Summer/Fall '98 I was working with Chris
Stauffer from AI Lab on VSAM
project, funded by DARPA Image Understanding program. We made a
system for detection car-person interactions. The system used Chris'
tracker and my stochastic parser to do it. The main trick was coming up
with a way to describe parallel event streams in a sequential machine.
The system was shown at the Integrated Feasibility Demonstration at
Carnegie-Mellon University in October '98, and then, at the Image
Understanding Workshop in November '98 in Monterrey, CA. There is a
CVPR'99 workshop paper (HTML)
about this work. Here
are the slides from my talk at CVPR'99. And another
paper with more technicalities. |
|
| Action recognition with SCFG |
|
|
For my Masters Thesis, I wrote a Stochastic
Context Free Grammar Parser which was initially developed by Andreas
Stolcke for NLP, and played with it a bit to make it
useable for vision. A toy problem I chose was musical conducting. The
technical chapters of the thesis are available as two separate papers TR
#449 (CVPR98, HTML),
and TR #450. |
|
| Lighting Independent Background Subtraction |
|
|
In November 1997, Claudio
Pinhanez staged a computer play, called It/I.
He wanted the play to have a truly theatrical feel, which required more
freedom with the stage lighting than machine vision could really handle.
We came up with a cute and simple way to cheat
lighting. It worked pretty well and even handled people using
flash photography during the performance you can see the movies on It/I's
web page and the papers below (or
see an HTML
version). |
|
| KidsRoom(I) |
|
|
KidsRoom
was the first "shocker" project that Aaron
put me on once I arrived at MIT in '96.
It gave me a clear preview of what's to come right off the bat :-). It
was a lot of work and a lot of fun and in retrospect I wish I had more
of a clue when I started. Stephen
and Jim did the wast
majority of work on this project. I only did the animation part. Unfortunately, it was a pretty large installation and
the only thing that remained from it is the 120M worth of web
pages and a paper. The next
version, KidsRoom2 was built by NearLife and shown at the Milenium Dome
in London for a year. |