Reproduce “Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior” 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
Zi-Ming Wang, Kai-Wen Duan, Kowei Huang, Akihiro Sugimoto, Shang-Hong Lai
The authors introduce a training method for 3D Gaussian Splatting that uses an L2 reconstruction loss and a Polarized Opacity Prior (POP) to strengthen Gaussians that dominate pixels and suppress others during training. On Tanks and Temples, it trains in 208 seconds with 269K Gaussians, versus 305 seconds and 318K for Taming-3DGS, with slightly lower PSNR: 23.618 versus 23.714 dB.
Key insights
- "Gradient leakage" is a primary reason low-opacity Gaussians persist, necessitating a more robust mechanism to manage their contribution.
- An L2 reconstruction loss, with its error-proportional gradients, is essential for stabilizing optimization and suppressing artifacts when applying aggressive opacity regularization.
- A Polarized Opacity Prior (POP) can effectively modulate training signals to drive important Gaussians towards full opacity and uninformative ones towards transparency, enabling continuous pruning and mitigating leakage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.