Reproduce “Compact 3D Scene Representation via Self-Organizing Gaussian Grids” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wieland Morgenstern, Florian Barthel, Anna Hilsmann, Peter Eisert
The authors compress 3D Gaussian Splatting scenes by sorting splat attributes into smooth, image-like grids and regularizing them during training before applying image compression. A shared ordering keeps each Gaussian’s linked attributes together for rendering. On Tanks & Temples, the file shrinks from 454 MB to 21.4 MB; PSNR is slightly higher, while LPIPS is slightly worse, showing similar reported quality rather than improvement on every metric.
Key insights
- Actively guiding the 3DGS optimization process towards a compressible representation during training is significantly more effective than applying post-hoc compression after vanilla training.
- Structuring high-dimensional Gaussian parameters into perceptually smooth 2D grids allows for drastic storage reduction using standard 2D image compression techniques.
- The inherent structure and regularization encourage the 3DGS optimization to converge with fewer Gaussians, leading to unexpected improvements in rendering speed without increasing training time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.