Graphical models for inference and learning in computer vision

Abstract

Graphical models are indispensable as tools for inference in computer vision, where highly structured and interdependent output spaces can be described in terms of low-order, local relationships. One such problem is that of graph matching, where the goal is to localise various parts of an object within an image: although the number of joint configurations of these parts may be very large, the relationships between them can typically be described in terms of simple skeletal structures, which lead to tractable inference.

We study problems of this type from three perspectives: firstly, how can we design graphical models to accurately model matching problems in computer vision? Secondly, how can machine learning be applied to such models, so that our inference algorithms can leverage the characteristics of a certain dataset? Thirdly, how can we exploit the specific energies that arise in this domain to develop faster inference algorithms, beyond the pessimistic worst-case results known for inference in graphical models? Naturally, these three problems are closely related, since both accurate and efficient inference are required by many learning algorithms.