Reproduce “On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy” 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
Letian Huang, Jiayang Bai, Jie Guo, Yuanqi Li, Yanwen Guo
The authors change the projection plane for each 3D Gaussian, using a tangent plane aligned with its direction from the camera rather than a shared image plane. This targets error from local affine approximation to perspective projection. At 0.2 times the original focal length, mean PSNR is 20.46 versus 15.58 for 3D-GS across six scenes, measured on matched central crops.
Key insights
- The inherent projection error in 3D-GS is minimized when the Gaussian's mean position is aligned with the camera's Z-axis, escalating rapidly as it deviates from this central axis or with wider fields of view.
- Minimizing projection errors can be achieved by performing the local affine approximation on an optimally chosen tangent plane specific to each Gaussian, rather than a single fixed plane for all Gaussians.
- Decoupling camera model specifics from the core Gaussian projection logic through unit sphere-based rasterization provides a unified and adaptable framework for various camera types.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.