Research map / 3D vision / Gaussian splatting
VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction
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.
Scales 3D Gaussian Splatting to large scenes by partitioning into cells with airspace-aware visibility-based data selection, parallel per-cell optimization, and a decoupled appearance module discarded at inference.
Problem
Scaling 3D Gaussian Splatting to large scenes is hampered by GPU memory limits, long optimization times, and appearance variations (exposure, illumination) that cause floaters, while NeRF-based large-scene methods lack detail and render slowly.
Approach
Divide-and-conquer: partition cameras and point cloud into grid cells, add extra training cameras via an airspace-aware visibility criterion (projected cell bounding box) and extra points via coverage, optimize cells in parallel, and merge; plus a decoupled appearance module that predicts a per-pixel transformation map from a downsampled render and per-image embedding, used only during training.
Key results
- On five large scenes VastGaussian achieves SSIM/LPIPS better than all compared methods (e.g., Campus SSIM 0.816, LPIPS 0.151 vs Modified 3DGS 0.712, 0.289), with PSNR comparable or better.
- Training takes about 2h25m-2h56m versus ~20-30h for NeRF baselines and Modified 3DGS, with VRAM 10.4-11.9G versus 31G for Modified 3DGS on a single GPU.
- Rendering reaches 126-172 FPS at 1080p, far above Mega-NeRF (0.25-0.31) and Switch-NeRF (0.12-0.15), though slower than Modified 3DGS (178-211).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add visibility-based camera selection (ablation) | VastGaussian · Sci-Art | SSIM: 0.694 -> 0.885 | Helped |
| Add coverage-based point selection (ablation) | VastGaussian · Sci-Art | SSIM: 0.874 -> 0.885 | Helped |
| Add visibility-based camera selection and coverage-based point selection (ablation) | VastGaussian · Sci-Art | SSIM: 0.699 -> 0.885 | Helped |
| Use airspace-aware rather than airspace-agnostic visibility (ablation) | VastGaussian · Sci-Art | SSIM: 0.855 -> 0.885 | Helped |
| Add decoupled appearance modeling (ablation) | VastGaussian · Sci-Art | SSIM: 0.858 -> 0.885 | Helped |
| Increase the number of cells from 8 to 16 or 24 (ablation) | VastGaussian · Sci-Art | SSIM: 0.885 -> 0.888; 0.885 -> 0.892 | Helped |
Surprising findings
- Higher cell counts beyond 8-16 yield only marginal quality gains while PSNR slightly decreases due to gradual brightness changes across distant cells.
- Lower PSNR on some scenes is attributed to over/under-exposed test images, while visual quality is judged better than the baseline.
Limitations
- Does not provide an optimal spatial division accounting for scene layout, cell number, and camera distribution.
- Very large scenes produce many Gaussians, increasing storage and slowing rendering.
- Quality metrics for Grid-NeRF are copied from its paper because its code lacks configs and images.
Experimental setup
Datasets: Mill-19 (Rubble, Building) UrbanScene3D (Campus, Residence, Sci-Art)
Benchmarks: Mill-19 UrbanScene3D
Metrics: SSIM PSNR LPIPS training time VRAM FPS at 1080p
Baselines: Mega-NeRF Switch-NeRF Grid-NeRF Modified 3DGS
Builds on: 3D Gaussian Splatting NeRF-W Mega-NeRF Block-NeRF Mip-NeRF 360
Closest papers by meaning
- CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes (ICLR 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS (NeurIPS 2024 · similarity 0.92)
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (CVPR 2025 · similarity 0.92)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Holistic Large-Scale Scene Reconstruction via Mixed 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.