Statistical Models of Shape: Optimisation and Evaluation

The goal of image interpretation is to convert raw image data into me- ingful information. Images are often interpreted manually. In medicine, for example, a radiologist looks at a medical image, interprets it, and tra- lates the data into a clinically useful form. Manual image interpretation is, however, a time-consuming, error-prone, and subjective process that often requires specialist knowledge. Automated methods that promise fast and - jective image interpretation have therefore stirred up much interest and have become a signi?cant area of research activity. Early work on automated interpretation used low-level operations such as edge detection and region growing to label objects in images. These can p- ducereasonableresultsonsimpleimages,butthepresenceofnoise,occlusion, andstructuralcomplexity oftenleadstoerroneouslabelling. Furthermore,- belling an object is often only the ?rst step of the interpretation process. In order to perform higher-level analysis, a priori information must be incor- rated into the interpretation process. A convenient way of achieving this is to use a ?exible model to encode information such as the expected size, shape, appearance, and position of objects in an image. The use of ?exible models was popularized by the active contour model, or ‘snake’ [98]. A snake deforms so as to match image evidence (e.g., edges) whilst ensuring that it satis?es structural constraints. However, a snake lacks speci?city as it has little knowledge of the domain, limiting its value in image interpretation.

Statistical models of shape, learnt from a set of examples, are a widely-used tool in image interpretation and shape analysis. Integral to this learning process is the establishment of a dense groupwise correspondence across the set of training examples.

This book gives a comprehensive and up-to-date account of the optimisation approach to shape correspondence, and the question of evaluating the quality of the resulting model in the absence of ground-truth data. It begins with a complete account of the basics of statistical shape models, for both finite and infinite-dimensional representations of shape, and includes linear, non-linear, and kernel-based approaches to modelling distributions of shapes. The optimisation approach is then developed, with a detailed discussion of the various objective functions available for establishing correspondence, and a particular focus on the Minimum Description Length approach. Various methods for the manipulation of correspondence for shape curves and surfaces are dealt with in detail, including recent advances such as the application of fluid-based methods.

This complete and self-contained account of the subject area brings together results from a fifteen-year program of research and development. It includes proofs of many of the basic results, as well as mathematical appendices covering areas which may not be totally familiar to some readers. Comprehensive implementation details are also included, along with extensive pseudo-code for the main algorithms. Graduate students, researchers, teachers, and professionals involved in either the development or the usage of statistical shape models will find this an essential resource.