Conference article

Clustering data is a standard tool to reduce large data sets enabling real-time rendering. When applying a grid based clustering, one cell of a chosen grid becomes the representative for a cluster cell. Starting from a uniform grid in a projective coordinate system, we investigate a redistribution of points from and to neighboring cells. By utilizing this redistribution, the grid becomes implicitly curvilinear, adapting to the point cloud’s inhomogeneous geometry. Additionally to pure point locations, we enabled data fields to influence the clustering behaviour. The algorithm was implemented as a CPU and a GPU code. The GPU implementation uses GLSL compute shaders for fast evaluation and directly manipulates the data on the graphics hardware, which reduces memory transfers. Data sets stemming from engineering and astrophysical applications were used for benchmarking. Different parameters dependent on the geometric properties were investigated and performance was measured. The method turned out to reach interactivity for medium sized point clouds and still good performance for large point clouds. The grid based approach is fast, while being able to adapt to the point cloud geometry.

No keywords available