Research map / 3D vision / Gaussian splatting
Volumetrically Consistent 3D Gaussian Rasterization
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.
Replaces 3DGS's splatting-based alpha with analytically integrated 3D Gaussian transmittance inside the same rasterizer, improving opaque-surface fidelity for view synthesis and matching SOTA tomography with fewer points.
Problem
3D Gaussian Splatting makes approximations (linearized exponential, no self-occlusion, affine covariance linearization via splatting) to the volume rendering equation, which reduces physical accuracy and makes opaque surfaces hard to represent.
Approach
Compute each 3D Gaussian's alpha by analytically integrating its density along the camera ray (closed form via erf, giving alpha = 1 - exp(-kappa G(gamma d) sqrt(2pi) beta)) and plug this into 3DGS's tile rasterizer, with a density reparameterization that favors high density for small Gaussians.
Key results
- On Mip-NeRF360 the method reaches SSIM 0.813 and LPIPS 0.209 at 136 FPS, versus 3DGS-Slang at 0.813 / 0.222 / 159 FPS, and on Tanks&Temples it reaches SSIM 0.854 and LPIPS 0.167 versus 3DGS-Slang's 0.850 / 0.176.
- On sparse-view (25 view) tomography it reports 35.08 3D-PSNR and 0.922 3D-SSIM on synthetic data (vs R2-Gaussian 34.97 / 0.921) using 58.2K points versus 66.4K.
- On real tomography data it matches R2-Gaussian (34.98 vs 35.01 PSNR) with 67.4K points, slightly more than R2-Gaussian's 71K on the same dataset metric table, and trains slower (30m23s vs 15m57s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace splatting-based alpha with analytic 3D Gaussian integration | 3DGS · Mip-NeRF360 | LPIPS: 0.222 -> 0.209 | Helped |
| Replace splatting-based alpha with analytic 3D Gaussian integration | 3DGS · Tanks&Temples | LPIPS: 0.176 -> 0.167 | Helped |
| Replace splatting-based alpha with analytic 3D Gaussian integration | 3DGS · DeepBlending | LPIPS: 0.248 -> 0.247 | No clear effect |
| Replace splatting-based alpha with analytic 3D Gaussian integration | 3DGS · Mip-NeRF360 | PSNR: 27.52 -> 27.30 | Hurt |
| Replace splatting-based alpha with analytic 3D Gaussian integration | 3DGS · Tanks&Temples | PSNR: 23.64 -> 23.74 | Helped |
| Replace splatting-based alpha with analytic 3D Gaussian integration | 3DGS · DeepBlending | PSNR: 29.78 -> 29.72 | Hurt |
| Use analytic Gaussian transmittance instead of the R2-Gaussian rendering model | R2-Gaussian · Synthetic Dataset, 25 views | 3D-PSNR: 34.97 -> 35.08 | Helped |
| Use analytic Gaussian transmittance instead of the R2-Gaussian rendering model | R2-Gaussian · Real Dataset, 25 views | 3D-PSNR: 35.01 -> 34.98 | No clear effect |
Surprising findings
- On Mip-NeRF360 PSNR the method is slightly below 3DGS-Slang (27.30 vs 27.52), attributed to 3DGS's diffuse-texture bias favoring PSNR.
- Opaque artifacts appear under large camera calibration errors, where 3DGS degrades more gracefully into blur.
Limitations
- Per-tile sorting and no overlap handling still cause popping artifacts.
- More sensitive than 3DGS to poor camera calibration and distortion, producing opaque artifacts.
- Training is somewhat slower than 3DGS due to more complex forward/backward computation.
Experimental setup
Datasets: Mip-NeRF360 Tanks and Temples DeepBlending R2-Gaussian tomography datasets (synthetic and real)
Benchmarks: Mip-NeRF360 Tanks and Temples DeepBlending R2-Gaussian tomography synthetic and real datasets
Metrics: PSNR SSIM LPIPS FPS 3D-PSNR 3D-SSIM training time point count
Baselines: 3DGS GES EVER Zip-NeRF Condor et al. Don't Splat R2-Gaussian FDK SART ASD-POCS IntraTomo NAF SAX-NeRF
Builds on: 3D Gaussian Splatting EWA splatting R2-Gaussian Condor et al. Don't Splat
Closest papers by meaning
- R$^2$-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction (NeurIPS 2024 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.90)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.90)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.90)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
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.