Adaptive Sampling for Geometric Approximation

dc.contributor.advisorMount, David Men_US
dc.contributor.authorAbdelrazek, Ahmed Abdelkaderen_US
dc.contributor.departmentComputer Scienceen_US
dc.contributor.publisherDigital Repository at the University of Marylanden_US
dc.contributor.publisherUniversity of Maryland (College Park, Md.)en_US
dc.date.accessioned2020-09-24T05:30:10Z
dc.date.available2020-09-24T05:30:10Z
dc.date.issued2020en_US
dc.description.abstractGeometric approximation of multi-dimensional data sets is an essential algorithmic component for applications in machine learning, computer graphics, and scientific computing. This dissertation promotes an algorithmic sampling methodology for a number of fundamental approximation problems in computational geometry. For each problem, the proposed sampling technique is carefully adapted to the geometry of the input data and the functions to be approximated. In particular, we study proximity queries in spaces of constant dimension and mesh generation in 3D. We start with polytope membership queries, where query points are tested for inclusion in a convex polytope. Trading-off accuracy for efficiency, we tolerate one-sided errors for points within an epsilon-expansion of the polytope. We propose a sampling strategy for the placement of covering ellipsoids sensitive to the local shape of the polytope. The key insight is to realize the samples as Delone sets in the intrinsic Hilbert metric. Using this intrinsic formulation, we considerably simplify state-of-the-art techniques yielding an intuitive and optimal data structure. Next, we study nearest-neighbor queries which retrieve the most similar data point to a given query point. To accommodate more general measures of similarity, we consider non-Euclidean distances including convex distance functions and Bregman divergences. Again, we tolerate multiplicative errors retrieving any point no farther than (1+epsilon) times the distance to the nearest neighbor. We propose a sampling strategy sensitive to the local distribution of points and the gradient of the distance functions. Combined with a careful regularization of the distance minimizers, we obtain a generalized data structure that essentially matches state-of-the-art results specific to the Euclidean distance. Finally, we investigate the generation of Voronoi meshes, where a given domain is decomposed into Voronoi cells as desired for a number of important solvers in computational fluid dynamics. The challenge is to arrange the cells near the boundary to yield an accurate surface approximation without sacrificing quality. We propose a sampling algorithm for the placement of seeds to induce a boundary-conforming Voronoi mesh of the correct topology, with a careful treatment of sharp and non-manifold features. The proposed algorithm achieves significant quality improvements over state-of-the-art polyhedral meshing based on clipped Voronoi cells.en_US
dc.identifierhttps://doi.org/10.13016/ml6t-2nzv
dc.identifier.urihttp://hdl.handle.net/1903/26384
dc.language.isoenen_US
dc.subject.pqcontrolledComputer scienceen_US
dc.subject.pquncontrolledapproximationen_US
dc.subject.pquncontrolledcomputational geometryen_US
dc.subject.pquncontrolledmeshingen_US
dc.subject.pquncontrollednearest-neighbor searchen_US
dc.subject.pquncontrolledsamplingen_US
dc.subject.pquncontrolledVoronoien_US
dc.titleAdaptive Sampling for Geometric Approximationen_US
dc.typeDissertationen_US

Files

Original bundle
Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
Abdelrazek_umd_0117E_21082.pdf
Size:
11.41 MB
Format:
Adobe Portable Document Format