This thesis presents a pipeline for compressing point clouds into a representation with 3D gaussians that, instead of directly compressing geometry, re-expresses the scene for optimized high-quality rendering from arbitrary viewpoints. As the procedure for optimising Gaussians requires photographs of the scene with known camera poses, which are often unavailable for point clouds, we replace them with synthetic views of rendered point clouds. We evaluate four rendering techniques for generating such synthetic views: camera-facing disks; disks and isotropic 2D Gaussians oriented according to surface normals; and a novel approach, DSORAC, which fills gaps between points based on depth, density, and distance to the nearest coloured point. We test the approach on four content- and acquisition-diverse groups of data: airborne laser-scanned scenes, isolated objects, a synthetic point cloud, and terrestrial mobile scans of a city, allowing us to assess the generality of the approach. In addition to comparing rendering techniques, we conduct further study of the effect of the initialisation point cloud sample size used in 3DGS optimisation. The effect of spherical harmonics on quality and compression was evaluated as well. Results show that the achieved compression ratios range from about 13 to over 170 times relative to the original point cloud depending on size of initialisation sample and degree of spherical harmonics. We further show that progressively increasing the degree of spherical harmonics provides a more effective and visually more consistent mechanism for adjusting the level-of-detail than the established approach of reducing the number of points in the representation.
|
