CoSurfGS: 3D Surface Gaussian Splatting with Collaborative Distributed Learning for Large-scale Scene Reconstruction

3D Gaussian Splatting (3DGS) has demonstrated impressive performance in scene reconstruction. However, most existing GS-based surface reconstruction methods focus on 3D objects or scenes with limited scale. Directly applying these methods to large-scale scene reconstruction will pose challenges such as high memory costs, excessive time consumption, and lack of geometric detail, which makes it difficult to implement in practical applications. 3D Gaussian Splatting (3DGS) has demonstrated impressive performance in scene reconstruction. However, most existing GS-based surface reconstruction methods are limited to single-machine settings and focus on individual objects or scenes with limited scale. When extended to large-scale scene reconstruction, these methods suffer from high memory consumption, prolonged training time, and insufficient geometric detail, which makes it difficult to implement in practical applications. To overcome these limitations, a natural solution is to distribute the learning workload across multiple collaborative devices. However, collaborative distributed learning presents unique challenges, including efficiently training models on resource-limited local devices and effectively integrating knowledge across devices to maintain global consistency and high-quality reconstruction. To address these challenges, we propose CoSurfGS, a novel collaborative-distributed framework that enables high-quality large-scale surface reconstruction while maintaining efficient training and GPU memory utilization. Specifically, we propose two modules: Local Model Compression (LMC), which eliminates redundant Gaussians to improve memory and training efficiency on each device; and Model Aggregation Schemes (MAS), which enhance global reconstruction quality by collaboratively distilling knowledge from multiple distributed devices. Extensive experiments on Urban3d, MegaNeRF, and BlendedMVS demonstrate that our proposed method can achieve fast and scalable high-fidelity surface reconstruction and photorealistic rendering.