Reproduce “GaussianPOP: Principled Simplification Framework for Compact 3D Gaussian Splatting via Error Quantification” 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
Soonbin Lee, Yeong-Gyu Kim, Simon Sasse, Tomas M. Borges, Yago Sanchez, Eun-Seok Ryu, Thomas Schierl, Cornelius Hellge
GaussianPOP prunes 3D Gaussian Splatting scenes by estimating rendered-color change from removing each Gaussian, as a single-Gaussian counterfactual under the current render. It supports pruning during training and on pre-trained models, recalculating scores between post-training rounds. On Bicycle at 80% pruning, PSNR is 20.71 versus LightGaussian’s 18.46 dB before fine-tuning and 25.11 versus 24.35 dB after 5,000 iterations, in this tested setting.
Key insights
- Directly quantifying the specific pixel error from a Gaussian's removal, rather than using heuristic importance scores, enables more accurate and visually coherent model simplification.
- The per-pixel error induced by removing a Gaussian can be analytically expressed as the squared difference between its color and the background it obscures, scaled by its effective contribution.
- Iterative re-quantification of Gaussian errors in post-training scenarios significantly improves pruning stability and accuracy by accounting for changes in transmittance as Gaussians are progressively removed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.