By Prof. Brian Potetz
The University of Kansas
Statistical approaches to computer vision problems are quickly becoming more feasible and more popular. Often, visual problems are modeled using Markov Random Fields or Factor Graphs where each variable node encodes a single pixel or image region. This results in highly interconnected graphs which make both inference and learning more difficult. In many cases, variables in these graphical models are real-valued or have many possible states, further complicating inference.
In this tutorial we will first present an overview of statistical models of natural images, and then discuss some statistical approaches to common computer vision applications. We will then discuss the particular approaches to inference (belief propagation, graph cuts) and learning (contrastive divergence, iterative scaling) that are commonly employed in the computer vision community.