| In this dissertation, a generic framework for the creation of hierarchical data representations for applications in scientific visualization based was developed using clustering. The major strength of the clustering-based approach for the construction of hierarchical representations for very large scientific data sets is its universal applicability. By using different error metrics, in combination with altering parameters of the clustering process, the clustering method can be used to solve a wide range of problems in scientific visualization and hierarchical data exploration. Adaptations of the general clustering framework were developed and used for the specific purposes of (1) feature extraction from volumetric data sets, (2) surface reconstruction from digitized/scanned objects, (3) simplification/coarsening of surface meshes, (4) simplification of vector field data in three-dimensional space, and (5) simplification of scalar field data in three-dimensional space.; The clustering framework supports the construction of data hierarchies for all commonly encountered discrete data formats used throughout scientific and engineering applications. This includes data sets given simply as a set of points without any connectivity information (“scattered Data”) as well as data defined on classical finite element/finite difference meshes (rectilinear meshes, curvilinear meshes, unstructured meshes, triangular and tetrahedral meshes, etc.).; In addition to the development of the framework for hierarchy construction via clustering, a scalable parallelization of the clustering approach was developed. The parallel implementation supports the analysis of very large data sets in greatly reduced response times on multi-processor machines and heterogeneous networks of workstations.; This dissertation demonstrated how the output of the clustering process, the “cluster tree,” can be used for view-dependent and adaptive rendering. View-dependent and adaptive rendering is concerned with the utilization of different data resolution levels in different regions of a data set to be visualized. As a consequence, one has to access nodes from different levels of the cluster tree to enable rendering at different resolutions in different regions in space. By coupling the view-dependent/adaptive paradigm with a scalable implementation of the hierarchy-building process, it is possible to create and interactively explore very large scientific data hierarchies in greatly reduced response times. |
