Research map / 3D vision / Gaussian splatting
GaussianPro: 3D Gaussian Splatting with Progressive Propagation
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 GaussianPro, which adds a PatchMatch-style progressive propagation of rendered depth/normal maps to 3DGS densification plus a planar normal/scale loss, improving novel view synthesis especially on textureless street scenes.
Problem
3DGS depends on SfM point-cloud initialization that is sparse on textureless surfaces in large-scale scenes, and its split/clone densification adds Gaussians without geometric priors, causing noisy geometry and poor novel-view rendering.
Approach
Render depth and normal maps from the Gaussians, propagate depth/normal candidates between neighboring pixels with PatchMatch and NCC photometric scoring, filter with multi-view geometric consistency, and back-project pixels whose propagated depth disagrees with rendered depth as new Gaussians; a planar loss aligns Gaussian normals with propagated normals.
Key results
- On Waymo, GaussianPro reaches 34.68 PSNR, 0.949 SSIM, 0.191 LPIPS, a 1.15 dB PSNR gain over 3DGS (33.53).
- On MipNeRF360, GaussianPro achieves 27.92 PSNR vs 27.88 for retrained 3DGS, a comparable result with slight improvement.
- Ablation on Waymo: propagation alone gives 34.48 PSNR and planar loss alone gives 34.02, versus 33.53 baseline and 34.68 for both.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add progressive propagation to 3DGS (ablation) | 3DGS · Waymo | PSNR: 33.53 -> 34.48 | Helped |
| Add planar constraint to 3DGS (ablation) | 3DGS · Waymo | PSNR: 33.53 -> 34.02 | Helped |
| Add planar constraint to propagation-based 3DGS (ablation) | Propagation only · Waymo | PSNR: 34.48 -> 34.68 | Helped |
| Add propagation to 3DGS with planar constraint (ablation) | Planar constraint only · Waymo | PSNR: 34.02 -> 34.68 | Helped |
| Use GaussianPro instead of 3DGS | 3DGS · Waymo | PSNR: 33.53 -> 34.68; 1.15 dB | Helped |
| Use GaussianPro instead of retrained 3DGS | 3DGS (Retrained) · MipNeRF360 | PSNR: 27.88 -> 27.92 | No clear effect |
| Use GaussianPro instead of 3DGS | 3DGS · MipNeRF360 room (30% training views) | PSNR: 28.45 -> 28.64 | Helped |
| Use GaussianPro instead of 3DGS | 3DGS · MipNeRF360 room (50% training views) | PSNR: 29.97 -> 30.27 | Helped |
| Use GaussianPro instead of 3DGS | 3DGS · MipNeRF360 room (70% training views) | PSNR: 30.87 -> 30.93 | Helped |
| Use GaussianPro instead of 3DGS | 3DGS · MipNeRF360 room (100% training views) | PSNR: 31.71 -> 31.98 | Helped |
| Initialize 3DGS with dense MVS points instead of SfM points | 3DGS · Street | PSNR: 35.05 -> 36.13 | Helped |
| Initialize 3DGS with dense MVS points instead of SfM points | 3DGS · Room | PSNR: 31.71 -> 32.05 | Helped |
Surprising findings
- Directly feeding denser MVS point clouds to 3DGS gives comparable or slightly better PSNR on the Street and Room scenes, but at roughly 4x training time and more Gaussians.
Limitations
- Does not specifically model dynamic objects, which produce artifacts as with other static Gaussian methods.
- Gains on MipNeRF360 are marginal, since its textured scenes are already well served by SfM initialization.
Experimental setup
Datasets: Waymo MipNeRF360
Benchmarks: Waymo MipNeRF360
Metrics: PSNR SSIM LPIPS FPS
Baselines: 3DGS Instant-NGP Mip-NeRF 360 Zip-NeRF
Builds on: 3D Gaussian Splatting ACMH NeuSG GaussianShader
Closest papers by meaning
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.93)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
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.