Research map / 3D vision / Gaussian splatting
3D-HGS: 3D Half-Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes 3D Half-Gaussian kernels, a plug-and-play replacement for 3D Gaussian kernels in splatting that uses two opacities per split Gaussian to better model discontinuities, improving novel view synthesis across several 3D-GS variants.
Problem
3D Gaussian Splatting uses Gaussian kernels that struggle to model shape and color discontinuities (sharp edges, object boundaries, texture), producing Gibbs-like overshoot and blurry edges in novel view synthesis.
Approach
Split each 3D Gaussian by a plane through its center into two halves with independent opacities (plane normal reuses the unused normal parameters), giving a closed-form integrated kernel and a rasterizer that computes per-half valid screen regions, so the kernel adds only one parameter per Gaussian.
Key results
- 3D-HGS improves 3D-GS PSNR on Mip-NeRF360 from 28.88 to 29.66 and on Tanks&Temples from 23.60 to 24.45.
- Mip-Splatting-HGS reaches 29.88 PSNR on Mip-NeRF360 and HGS-MCMC reaches 30.13, with HGS-MCMC setting a new state of the art on that dataset.
- Half-Gaussian kernels keep rendering speed and memory close to baselines; 3D-HGS runs at 125 FPS vs 115 FPS for 3D-GS on Mip-NeRF360.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Half-Gaussian kernels | 3D-GS · Mip-NeRF360 | PSNR: +0.78 | Helped |
| Add Half-Gaussian kernels | 3D-GS · Tanks&Temples | PSNR: +0.85 | Helped |
| Add Half-Gaussian kernels | 3D-GS · Deep Blending | PSNR: +0.35 | Helped |
| Add Half-Gaussian kernels | Scaffold-GS · Mip-NeRF360 | PSNR: +0.30 | Helped |
| Add Half-Gaussian kernels | Scaffold-GS · Tanks&Temples | PSNR: +0.46 | Helped |
| Add Half-Gaussian kernels | Scaffold-GS · Deep Blending | PSNR: +0.15 | Helped |
| Add Half-Gaussian kernels | Mip-Splatting · Mip-NeRF360 | PSNR: +0.49 | Helped |
| Add Half-Gaussian kernels | Mip-Splatting · Tanks&Temples | PSNR: +0.78 | Helped |
| Add Half-Gaussian kernels | Mip-Splatting · Deep Blending | PSNR: +0.15 | Helped |
| Add Half-Gaussian kernels | GS-MCMC · Mip-NeRF360 | PSNR: +0.24 | Helped |
| Add Half-Gaussian kernels | GS-MCMC · Tanks&Temples | PSNR: +0.77 | Helped |
| Add Half-Gaussian kernels | GS-MCMC · Deep Blending | PSNR: +0.13 | Helped |
Surprising findings
- Scaffold-HGS gains are smaller on Mip-NeRF360 (+0.30) and it reports a slight SSIM and LPIPS trade-off in some settings, while HGS-MCMC shows lower SSIM on Tanks&Temples (0.841 vs 0.860).
Limitations
- Gains on some metrics are mixed; e.g., Mip-Splatting-HGS and HGS-MCMC lower SSIM on some datasets, and Deep Blending gains are small (0.13-0.35 dB).
- Experiments are limited to four baseline Gaussian splatting methods and three datasets, all run on one RTX 3090.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS FPS memory
Baselines: Mip-NeRF 2D-GS FreGS GES 3D-GS Scaffold-GS Mip-Splatting GS-MCMC
Builds on: 3D Gaussian Splatting Mip-Splatting Scaffold-GS GS-MCMC 2D Gaussian Splatting
Closest papers by meaning
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.93)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.92)
- Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.91)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.