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

TL;DR

This work tackles the memory and training-time bottlenecks of large-scale 3D Gaussian Splatting by introducing DOGS, a distributed training framework that recursively splits a scene into $K$ blocks and maintains a global Gaussian model on a master node. Local blocks train in parallel and are regularized to converge to the global model via an ADMM-based consensus, with adaptive penalties and over-relaxation to accelerate convergence. Empirically, DOGS achieves $6\times$ to $8\times$ faster training while delivering state-of-the-art novel view synthesis quality across Mill19, UrbanScene3D, and MatrixCity datasets. The method enables scalable city-scale 3D reconstruction with a single global model for inference, highlighting its practical impact for real-time rendering and large-scale applications.

Abstract

The recent advances in 3D Gaussian Splatting (3DGS) show promising results on the novel view synthesis (NVS) task. With its superior rendering performance and high-fidelity rendering quality, 3DGS is excelling at its previous NeRF counterparts. The most recent 3DGS method focuses either on improving the instability of rendering efficiency or reducing the model size. On the other hand, the training efficiency of 3DGS on large-scale scenes has not gained much attention. In this work, we propose DoGaussian, a method that trains 3DGS distributedly. Our method first decomposes a scene into K blocks and then introduces the Alternating Direction Method of Multipliers (ADMM) into the training procedure of 3DGS. During training, our DOGS maintains one global 3DGS model on the master node and K local 3DGS models on the slave nodes. The K local 3DGS models are dropped after training and we only query the global 3DGS model during inference. The training time is reduced by scene decomposition, and the training convergence and stability are guaranteed through the consensus on the shared 3D Gaussians. Our method accelerates the training of 3DGS by 6+ times when evaluated on large-scale scenes while concurrently achieving state-of-the-art rendering quality. Our code is publicly available at https://github.com/AIBluefisher/DOGS.

Figures (11)

• Figure 1: DoGaussian accelerates 3D GS training on large-scale scenes by $6+$ times with better rendering quality.
• Figure 2: The pipeline of our distributed 3D Gaussian Splatting method. 1) We first split the scene into $K$ blocks with similar sizes. Each block is extended to a larger size to construct overlapping parts. 2) Subsequently, we assign views and points into different blocks. The shared local 3D Gaussians (connected by solid lines in the figure) are a copy of the global 3D Gaussians. 3) The local 3D Gaussians are then collected and averaged to the global 3D Gaussians in each consensus step, and the global 3D Gaussians are shared with each block before training all blocks. 4) Finally, we use the final global 3D Gaussians to synthesize novel views.
• Figure 3: Scene splitting results of our method v.s. VastGaussian lin2024vastgaussian. (a) VastGaussian can result in imbalanced blocks. (b) Our recursive bipartite strategy solves the imbalanced splitting issue. (c) Points and views with the same grid coordinate are assigned to the same block.
• Figure 4: Qualitative comparisons of our method and others on the Mill19 dataset. The first row and second row are respectively the results of scene 'building' and 'rubble'.
• Figure 5: Qualitative comparisons of our method and others on the UrbanScene3D dataset. From top to bottom are respectively the results of scenes 'campus', 'residence', and 'sci-art'.