• Home
• About
• People
• News
• Events
• Research
• Courses
• Graduate Study
• Undergraduate Study
• Places
• Computing Systems
• Industrial Relations
• Alumni
• Contacts

Brown CS: Computer Vision and Learning Seminar Series

Fall 2004

Upcoming Talks

TBA

Spring 2004

Past talks

Friday, February 20, 2004 at 3:00pm - Lubrano Conference Room:

Nonparametric Belief Propagation  
Erik B. Sudderth and Alexander T. Ihler, MIT AI Lab

Friday, February 6, 2004 at 3:00pm - Lubrano Conference Room:

Using the forest to see the trees: a graphical model relating features, objects and scenes  
Kevin Murphy, MIT AI Lab

Talk Details

Detecting objects, such as faces or cars, in images is an important but difficult problem. Most current approaches try to classify each local image patch in isolation, disregarding the rest of the image. Such local approaches may suffer from inherent ambiguities. In addition, it is computationally expensive to find the right local regions to classify, often requiring exhaustive search in position and scale. We show how using a low-dimensional summary of the whole image (the "gist" of the image) can ameliorate both problems, since the gist can provide a prior on what kinds of objects to expect, and where to expect them. The gist can also be used to recognize "scenes", or correlated patterns of object co-occurrences. We model the

trained graphical model, which combines global and local information.

(Joint work with Antonio Torralba and Bill Freeman)

Bio

Kevin Murphy received his PhD from UC Berkeley in 2002, where he worked with Stuart Russell and Michael Jordan. His thesis was called 'Dynamic Bayesian Networks: Representation, Inference and Learning'. He is currently a post-doctoral researcher at the MIT AI Lab, where he works with Leslie Kaelbling and Bill Freeman on applying graphical models and machine learning to problems in computer vision and mobile robotics.

Graphical models provide a powerful general framework for formulating and solving problems of statistical inference and machine learning. In many applications of graphical models, the hidden variables of interest are most naturally specified by continuous, non-Gaussian distributions. However, due to the limitations of existing inference algorithms, it is often necessary to form coarse, discrete approximations to such models.

In this talk, we describe a nonparametric belief propagation (NBP) algorithm, which uses stochastic methods to propagate kernel-based approximations to the true continuous messages. Each NBP message update requires approximating the product of several Gaussian mixtures; we present efficient procedures for sampling from this product using multiscale representations. We demonstrate NBP's effectiveness on two different applications: visual recognition and tracking of complex objects, and distributed self-localization of an ad hoc sensor network.