Level of Detail Generation for Point Clouds
With the increasing quality of 3D models the conventional approach of representing 3D models as a closed polygonal surfaces looses efficiency. Point Clouds offer an alternative by only using point samples of the 3D surface. These samples do not require connectivity information, drastically simplifying the rendering process. In addition, data from 3D Scans, LIDAR, etc. is often only available as point data, and would otherwise have to be converted into a polygonal model. Point clouds can often consist of hundreds of millions or even billions of samples. It is therefore necessary to find a way to simplify regions that are further away from the camera in order to not draw an entire dataset every frame. This is commonly known as Level of Detail. This Thesis revisits the sub sampling approach presented in Potree. Afterwards we explore clustering based simplification techniques as a way to not only use the position of samples, but also take advantage additional information such as curvature and color. Human-made objects often contain planar surfaces. By using geometric information as a criterion for creating our level of detail hierarchy, we manage to drastically reduce the sample density in such planar regions, while still being able to conserve complex geometries.
Advisor: Henrik Masbruch
Supervisor: Prof. Dr. Rüdiger Westermann