# DoGaussian: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus

**Authors:** Yu Chen, Gim Hee Lee  
**arXiv:** [2405.13943](https://arxiv.org/abs/2405.13943)  
**Date:** 2024-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.13943v1/

## Key points

- Recent advancements in 3D Gaussian Splatting (3DGS) have shown promising outcomes in novel view synthesis (NVS)
- DoGaussian method introduced to address training efficiency for large-scale scenes by breaking down scenes into blocks and using ADMM
- DoGaussian maintains a global 3DGS model on the master node and K local models on slave nodes during training, leading to faster convergence
- Significantly reduces training time by over six times without compromising rendering quality on large-scale scenes
- Traditional photogrammetry methods like Structure-from-Motion (SfM) and keypoint extraction with SIFT are crucial for reconstructing sparse scene structures, especially for city-scale scenes.

## Summary

, , , , 

In the realm of 3D reconstruction, recent advancements in 3D Gaussian Splatting (3DGS) have showcased promising outcomes in the domain of novel view synthesis (NVS). With its exceptional rendering capabilities and high-fidelity output, 3DGS has outperformed its predecessors like NeRF. Previous iterations of 3DGS focused on enhancing rendering efficiency or reducing model size, but overlooked training efficiency for large-scale scenes. To address this gap, a new method called DoGaussian has been introduced. It involves breaking down a scene into K blocks and integrating the Alternating Direction Method of Multipliers (ADMM) into the training process. During training, DoGaussian maintains a global 3DGS model on the master node alongside K local models on slave nodes. Once training is complete, only the global model is used during inference. DoGaussian significantly reduces training time through scene decomposition while ensuring convergence and stability via consensus on shared 3D Gaussians. By leveraging this methodology, it accelerates training by over six times without compromising rendering quality on large-scale scenes. In the context of large-scale 3D reconstruction, traditional photogrammetry methods such as Structure-from-Motion (SfM) and keypoint extraction with SIFT are pivotal in reconstructing sparse scene structures. For city-scale scenes, a 'divide-and-conquer' strategy is commonly employed for enhanced extensibility and scalability within 3D reconstruction systems. , , , , 

Overall, with its innovative distributed-oriented approach and emphasis on large-scale scene reconstruction, DoGaussian represents a significant leap forward in advancing the capabilities of 3D Gaussian Splatting for cutting-edge applications in computer vision and graphics.