Abstract

Unsupervised and weakly supervised extraction of information from image and video data

Martial Hebert
Carnegie Mellon University
The Robotics Institute

This talk will review several formulations for classification over graphs. We describe how conditional field models can be used effectively
for parameter learning and inference. We compare the performance of different appraoches for learning and parameter estimation over the
field models. The application domain that we are considering is the use of context in image interpretation. This includes representating
the relations between images parts, between objects, or between regions. We show how these different types of relcations can be unified in a single conditional field and show results on a number of image databases and applications.

Presentation (PDF File)
Video of Talk (RealPlayer File)