Ole Siemers
Supervisor(s): Diana Marin
TU Wien
Abstract: This work presents a coarse-to-fine optimisation method for 3D Gaussian Splatting(3DGS) that constructs a Level of Detail (LoD) hierarchy during optimisation, which can be rendered selectively. By gradually adjusting the resolution, the method reduces computational effort, speeds up optimisation and generates a LoD hierarchy in the process. Based on the sampling rate, defined as the ratio of the resolution at which the model was optimised to that at which it is viewed, a selective rendering method is presented. Selective rendering reduces the number of primitives processed and mitigates aliasing errors, at the cost of increased memory usage on the Graphics Processing Unit (GPU) due to multiple independent LoD levels. The method is evaluated using 3DGS and Elliptical Weighted Average (EWA)-filtering as a basis for comparison on common 360° and aerial image datasets, with a focus on low-resolution renderings and distant viewpoints.The results show that the method speeds up optimisation and reduces the number of processed primitives. Particularly for distant or low-resolution views, images are generated more quickly, and aliasing errors are reduced. At full resolution, the visual quality remains approximately the same as the baseline. Although the method requires additional GPU memory during rendering, it offers a practical approach to faster optimisation of more compact models that are rendered with reduced aliasing.
Keywords: 3D Reconstruction, Point-based Graphics, Real-time Graphics, Rendering
Full text: Year: 2026
