Shape Analysis towards Model-based Segmentation

Abstract

In this chapter, we investigate shape modeling and registration towards model-based shape-driven object extraction and segmentation. The chapter consists of two major contributions: (i) a variational level set approach for global-to-local shape registration and (ii) an energetic formulation to impose prior shape knowledge within implicit representations. Distance transforms are proven to be a very efficient feature space to perform shape registration. Prior knowledge is an important tool in the segmentation process. Following the example of shape registration, we introduce a stochastic level set prior that constrains the segmentation result to be in a family of shapes defined through a similarity transformation of the prior model. Promising results and systematic validation for each of the topics under investigation demonstrate the performance and the potentials of our approach.