Research map / 3D vision / Gaussian splatting
DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus
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.
Distributed large-scale 3D Gaussian Splatting training via recursive scene splitting and ADMM-based Gaussian consensus, yielding a single global model that is queried at inference.
Problem
Training 3D Gaussian Splatting on large-scale city-scale scenes is slow and memory-hungry, and prior divide-and-conquer methods either need multiple sub-models at inference or lack guaranteed training convergence.
Approach
Split the scene into K overlapping blocks via recursive balanced bipartition, train local 3DGS models per block in parallel, and enforce consistency of shared Gaussians with ADMM consensus (global model averaged from local models every 100 iterations, with adaptive penalties and over-relaxation); only the global model is kept for rendering.
Key results
- On Mill19/UrbanScene3D, DOGS achieves the best PSNR in Building (22.73), Rubble (25.78), and Sci-Art (24.42 in the second-best row) while training in about 2-4.5 hours versus 18-23 hours for 3DGS, i.e. roughly 6-8x faster.
- On MatrixCity Small City, DOGS reaches the best rendering quality where 3DGS did not finish training within two days and VastGaussian failed.
- Ablation: removing the consensus step drops PSNR from 25.78 to 22.80 on the Campus-averaged setting, and removing adaptive penalties drops PSNR by about 1.5 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 3D Gaussian consensus (ablation) | DOGS distributed training · Rubble (Mill19) | PSNR: 22.80 -> 25.78 | Helped |
| add self-adaptation of penalty parameters (ablation) | DOGS distributed training · Rubble (Mill19) | PSNR: 24.30 -> 25.78 | Helped |
| add over-relaxation (ablation) | DOGS distributed training · Rubble (Mill19) | PSNR: 24.45 -> 25.78 | Helped |
| increase overlap scale factor from 1.2 to 1.3 (ablation) | DOGS · Rubble (Mill19) | PSNR: 24.25 -> 24.86 | Helped |
| increase overlap scale factor from 1.3 to 1.4 (ablation) | DOGS · Rubble (Mill19) | PSNR: 24.86 -> 25.78 | Helped |
| increase overlap scale factor from 1.4 to 1.5 (ablation) | DOGS · Rubble (Mill19) | PSNR: 25.78 -> 25.97 | Helped |
| use DOGS instead of the strongest listed baseline | 3DGS-based large-scale reconstruction · Building (Mill19) | PSNR: 22.53 -> 22.73 | Helped |
| use DOGS instead of the strongest listed baseline | 3DGS-based large-scale reconstruction · Rubble (Mill19) | PSNR: 25.51 -> 25.78 | Helped |
| use DOGS instead of the strongest listed baseline | 3DGS-based large-scale reconstruction · Campus (UrbanScene3D) | PSNR: 23.82 -> 24.01 | Helped |
| use DOGS instead of the strongest listed baseline | 3DGS-based large-scale reconstruction · Residence (UrbanScene3D) | PSNR: 22.57 -> 21.94 | Hurt |
| use DOGS instead of the strongest listed baseline | 3DGS-based large-scale reconstruction · Sci-Art (UrbanScene3D) | PSNR: 26.51 -> 24.42 | Hurt |
| use DOGS instead of the strongest listed baseline | 3D reconstruction · MatrixCity aerial | PSNR: 28.33 -> 28.58 | Helped |
Surprising findings
- VastGaussian trains faster than DOGS on the benchmarks because DOGS incurs data-transmission overhead, yet DOGS renders higher quality.
- Removing the ADMM consensus step causes a large quality drop, showing that naive distributed block training produces block-boundary artifacts.
Limitations
- Additional communication overhead and reliance on a master node make implementation less flexible than VastGaussian.
- GPU memory remains a constraint when zooming out to larger areas with more Gaussians.
- Scene splitting overlap constant is chosen empirically without theoretical justification; convergence proof uses fixed penalty only up to 2000 iterations.
Experimental setup
Datasets: Mill19 UrbanScene3D MatrixCity
Benchmarks: Mill19 UrbanScene3D MatrixCity
Metrics: PSNR SSIM LPIPS training time FPS memory number of points
Baselines: Mega-NeRF Switch-NeRF 3DGS Fed3DGS VastGaussian Hierarchy-GS
Builds on: 3D Gaussian Splatting ADMM VastGaussian Mega-NeRF
Closest papers by meaning
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.92)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · 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.