Research map / 3D vision / Gaussian splatting
Holistic Large-Scale Scene Reconstruction via Mixed 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 MixGS, a holistic large-scale 3DGS framework that decodes view-aware hash-encoded features of frustum-visible Gaussians into extra Gaussians and mixes them with the coarse originals, training on a single 24GB GPU.
Problem
Divide-and-conquer 3DGS methods for large-scale scene reconstruction need complex partition/threshold tuning and lose global scene consistency across blocks.
Approach
Treat the whole scene holistically: encode view-frustum Gaussian positions with multi-resolution hash encoding, modulate them with auxiliary Gaussian attributes and camera pose, decode residual Gaussians via a multi-head MLP and offset pool, and mix them with the original visible Gaussians for rendering, implicitly replacing explicit densification.
Key results
- MixGS achieves 23.39 PSNR on Residence and the highest SSIM on all four scenes at 4x downsampling (e.g., 0.815 on Residence).
- On Rubble, MixGS reaches 26.66 PSNR versus 25.78 for the next best (DOGS).
- MixGS renders at 33.5-53.9 FPS on a single RTX 3090 while training on a single 24GB GPU, versus <0.1 FPS for NeRF baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-resolution hash encoding (ablation) | MixGS · Rubble test views | PSNR: 24.70 -> 26.66 | Helped |
| Add auxiliary Gaussian enhancement (ablation) | MixGS · Rubble test views | PSNR: 26.24 -> 26.66 | Helped |
| Add the coarse Gaussian training stage (ablation) | MixGS · Rubble test views | PSNR: 25.86 -> 26.66 | Helped |
| Add the implicit feature training stage (ablation) | MixGS · Rubble test views | PSNR: 25.99 -> 26.66 | Helped |
| Use a learned offset pool instead of direct position regression (ablation) | MixGS · Rubble test views | PSNR: 26.35 -> 26.66 | Helped |
| Mix view-frustum Gaussians with decoded Gaussians (ablation) | MixGS · Rubble test views | PSNR: 23.65 -> 26.66 | Helped |
| Use MixGS instead of the strongest listed baseline | Large-scale scene reconstruction · Building test views | SSIM: 0.764 -> 0.771 | Helped |
| Use MixGS instead of the strongest listed baseline | Large-scale scene reconstruction · Rubble test views | SSIM: 0.785 -> 0.792 | Helped |
| Use MixGS instead of the strongest listed baseline | Large-scale scene reconstruction · Residence test views | SSIM: 0.805 -> 0.815 | Helped |
| Use MixGS instead of the strongest listed baseline | Large-scale scene reconstruction · Sci-Art test views | SSIM: 0.833 -> 0.856 | Helped |
| Increase image downsampling from 4× to 6× | MixGS · Building test views | PSNR: 23.03 -> 23.82 | Helped |
| Increase image downsampling from 4× to 6× | MixGS · Rubble test views | PSNR: 26.66 -> 27.95 | Helped |
Surprising findings
- NeRF-based methods (Switch-NeRF) obtain higher PSNR than 3DGS methods on Sci-Art, which the authors attribute to the dataset's inherent blur.
Limitations
- Training time is longer than multi-GPU divide-and-conquer approaches due to single-GPU setup.
- Gains in PSNR/LPIPS are not uniformly best; LPIPS on Building and Rubble trails DOGS/VastGaussian.
- Evaluated only on four scenes from two drone datasets.
Experimental setup
Datasets: UrbanScene3D Mill19
Benchmarks: Mill19 Building Mill19 Rubble UrbanScene3D Residence UrbanScene3D Sci-Art
Metrics: PSNR SSIM LPIPS FPS
Baselines: Mega-NeRF Switch-NeRF GP-NeRF 3DGS VastGaussian Hierarchy-GS DOGS CityGaussian
Builds on: 3D Gaussian Splatting multi-resolution hash encoding Scaffold-GS Mega-NeRF
Closest papers by meaning
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.91)
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction (CVPR 2024 · similarity 0.90)
- On Scaling Up 3D Gaussian Splatting Training (ICLR 2025 · similarity 0.90)
- Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.90)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 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.