UC San Diego

Meshes generated from a typical manual segmentation process are often unsuitable for simulation purposes due to poor element quality and artifacts. In this thesis we describe a remeshing approach for converting low quality triangular meshes into spatially realistic, simulation quality meshes. We improve upon the Laplacian Mesh Optimization remeshing framework by incorporating topology modification, triangle area equalization, and feature detection using computer vision techniques. We train a supervised edge learning algorithm with human produced contours and corresponding image data. The resulting classifier is used to generate a probabilistic edge map for the entire image. Salient features in the mesh surface are detected as a weighted combination of surface curvature and the learned image boundary probability. We compare the remeshing performance of our algorithm with the original method on real world data sets, showing that our approach can produce higher quality meshes from extremely irregular input meshes while simultaneously enhancing spatial realism