Computer Science Colloquium

Probabilistic graphical models for scene and object recognition

Kevin Murphy
MIT CSAIL (Computer Science & Artificial Intelligence Laboratory)

Friday, April 23, 2004 11:30 A.M.
Room 1302 Warren Weaver Hall
251 Mercer Street
New York, NY 10012-1185

Directions: http://cs.nyu.edu/csweb/Location/directions.html
Colloquium Information: http://cs.nyu.edu/csweb/Calendar/colloquium/index.html

Hosts:

Richard Cole cole@cs.nyu.edu, (212) 998-3119

Abstract

Probabilistic graphical models are a way of combining multiple sources of noisy evidence together in a principled fashion, in order to come up with an optimal estimate of the hidden state of a system. Well-known examples include Kalman filters and HMMs. In this talk, I will show how we can use graphical models to perform fast and robust place and scene recognition. I will then show how to extend the model to detect objects such as cars, people, computers, etc. We use the output of the scene recognition system to decide which objects are likely to be present (for example, cars are unlikely in indoor scenes). Next we use global image features to predict the likely location of the object. Finally we apply a standard object detector (based on boosted decision stumps) to the image. The various sources of information are combined using a discriminatively trained graphical model (a conditional random field). We discuss some methods for efficiently training such models, and demonstrate our system on a challenging dataset of indoor and outdoor images collected with a wearable camera.