PDF version of this document

Project Background and Description

R. S. Schestowitz
Imaging Science and Biomedical Engineering
Stopford Building
University of Manchester
United Kingdom

Abstract:

Landmark selection in appearance models proves to be a complex problem that is presently solved in ways which are by no obvious criterion optimal. Given a set of training data, diffeomorphic warps are used to pose a simpler correspondence problem, thereby generating more compact models of appearance. In recent work, statistical models of intensity and shape are combined in a rather artificial way that ignores the inherent correlation between the two. Furthermore, construction of appearance models has not yet been made a fully automated process. The need for manual annotation seems redundant given some intriguing developments in model-based image analysis and exploitations of these developments bears potential with respect to optimisation and complete automation. This paper attempts to describe the existing problems in some depth and outline some of the currently conceived solutions.