Reproduce “Gradient-Driven Natural Selection for Compact 3D Gaussian Splatting” 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
Xiaobin Deng, Qiuli Yu, Changyu Diao, Min Li, Duanqing Xu
Natural Selection prunes 3D Gaussian Splatting scenes by applying shared pressure to Gaussian opacities while rendering-quality gradients provide per-Gaussian feedback. It removes primitives whose opacity falls below 0.001. On Mip-NeRF 360, it reports 28.13 dB PSNR with 466,667 Gaussians versus 27.50 dB and 3,320,453 for 3DGS; its base is Improved-GS, and its regularization rate is tuned per scene.
Key insights
- Gaussian pruning can be effectively and adaptively managed by treating each Gaussian's opacity as its vitality and allowing it to compete against a uniform environmental pressure via its rendering gradient.
- Applying regularization to the pre-activation opacity parameter with a 'finite prior' selectively accelerates the removal of weak Gaussians while preserving strong ones, optimizing the balance between pruning efficiency and fairness.
- A self-regulating pruning mechanism can be achieved by utilizing only the intrinsic optimization gradients of the 3DGS system, avoiding external heuristics or architectural modifications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.