Speaker: Klaus, Philip (TU Wien)
Rendering arbitarily large point clouds requires level of detail (LOD) structures that allow users to stream and render representative subsets of the data in real-time. This thesis introduces PotreeConverterGpu, a GPU-based software for generating octree-based LOD structures from point clouds. It is based on PotreeConverter and investigates GPU-friendly algorithms to generate the same structure up to 85 times faster for similar quality. The thesis also explores color-filtering for lower levels of detail to improve the quality for point clouds in a fashion similar to mip mapping for textured meshes.