Level-set methods and geodesic distance functions

Sammanfattning: The work in this thesis focuses on efficient implementations of level-set methods and geodesic distance functions. The level-set method is a grid based design that inherits many favorable traits from implicit geometry. It is connected to distance functions through its special way of representing geometry: in ìo each point in space stores the closest distance to the surface. To differentiate between the inside and outside of a closed object a signed distance is used. In the discrete form the representation keeps a box around the surface that stores regularly positioned samples of the distance function – i.e. a grid. These samples implicitly encode the surface as the zeroth level-set of the signed distance function, hence the name level-set methods. With this representation of geometry follows a toolbox of operations based on partial differential equations (PDE). The solution to these PDES allows for arbitrary motion and deformation of the surface. This thesis focuses on two topics: 1) grid storage for level-set methods, and 2) geodesic distance functions and parameterization. These topics are covered in a series of in-depth articles. Today, level-set methods are becoming widespread in both academia and industry. Data structures and highly accurate methods and numerical schemes are available that allow for efficient handling of topological changes of dynamic curves and surfaces. For some applications, such as the capturing of the air/water interface in free surface fluid simulations, it’s is the only realistic choice. In other areas level-set methods are emerging as a competitive candidate to triangle meshes and other explicit representations. In particular this work introduces efficient level-set data-structures that allow for extremely detailed simulations and representations. It also presents a parameterization method based on geodesic distance that produces a unique coordinate system, the Riemannian normal coordinates (RNC). Amongst other interesting applications this parameterization can be used for decal compositing, and the translation of vector space algorithms to surfaces. The approximation of the RNC involves one or more distance functions. In this thesis, a method originally presented for triangle meshes is adopted. It is then and extended to compute accurate geodesic distance in anisotropic domains in two and three dimensions. The extension to higher dimensions is also outlined. To motivate this work several applications based on these novel methods and data structures are presented showing rapid ray-tracing, shape morphing, segmentation, geodesic interpolation, texture mapping, and more.