High-quality Approximation of Scientific Data using 3D Gaussian Splatting
Andres Sewell, Landon Dyken, Victor A. Mateevitsi, Will Usher, Jefferson Amstutz, Thomas Marrinan, Khairi Reda, Silvio Rizzi, Joseph A. lnsley, Michael E. Papka, Sidharth Kumar, Steve Petruzza · 2024
This work explores the application of recent advancements in radi-ance field rendering, specifically 3D Gaussian splatting, to generate high-quality approximations of scientific data. In this technique, a 3D Gaussian splatting model is built from a 3D point cloud gen-erated using structure-from-motion, or a randomly initialized one when using a NeRF as input. This point cloud serves as the basis for initializing a set of Gaussian primitives, which are then refined through machine learning to minimize differences between ground truth and rendered images. We modified this pipeline to train Gaus-sian models directly from scientific data, eliminating the need for structure-from-motion. We test exporting an isosurface as a point cloud, which is then used to train a Gaussian model representing the dataset's isosurface. We also experimented with using a cinema database to produce a 3D Gaussian model; however, this approach yielded less promising results due to sub-optimal point cloud ini-tialization. Our findings highlight the potential of this technique for scientific datasets, suggesting it could enable efficient post-hoc visualization with reduced computational resources.