GigaGS: Scaling up Planar-Based 3D Gaussians for Large Scene Surface Reconstruction

TL;DR

GigaGS tackles large-scale scene surface reconstruction with 3D Gaussian Splatting by introducing a mutual-visibility–driven partitioning scheme and a Level-of-Detail (LoD) framework to enable parallel optimization. It couples a hierarchical plane representation with anchor Gaussians, an octree-like partitioning strategy, and appearance/geometry regularization to preserve multi-view consistency across scales. The approach achieves high-quality surfaces and meshes on giga-scale scenes, leveraging TSDF fusion for final mesh extraction and demonstrating scalable Gaussian counts through partitioned optimization. This work enables accurate, scalable surface reconstruction for extensive environments and supports downstream applications in VR, navigation, and robotics.

Abstract

3D Gaussian Splatting (3DGS) has shown promising performance in novel view synthesis. Previous methods adapt it to obtaining surfaces of either individual 3D objects or within limited scenes. In this paper, we make the first attempt to tackle the challenging task of large-scale scene surface reconstruction. This task is particularly difficult due to the high GPU memory consumption, different levels of details for geometric representation, and noticeable inconsistencies in appearance. To this end, we propose GigaGS, the first work for high-quality surface reconstruction for large-scale scenes using 3DGS. GigaGS first applies a partitioning strategy based on the mutual visibility of spatial regions, which effectively grouping cameras for parallel processing. To enhance the quality of the surface, we also propose novel multi-view photometric and geometric consistency constraints based on Level-of-Detail representation. In doing so, our method can reconstruct detailed surface structures. Comprehensive experiments are conducted on various datasets. The consistent improvement demonstrates the superiority of GigaGS.

Figures (6)

• Figure 1: We propose GigaGS, the first work specifically designed for large scene surface reconstruction. Our approach ensures high rendering quality while also extracting high-quality meshes.
• Figure 2: Visualization of the effects at different levels. We visualized the rendered images and normal maps obtained by rendering different levels of the same scene as rendering entities.
• Figure 3: Different loss terms on the final optimization process. Initially, using only the image loss as supervision does not yield a satisfactory surface reconstruction. The incorporation of an appearance model reduces certain artifacts. However, the addition of flatten regularization on the geometric structure without additional geometric supervision leads to a decrease in expressiveness by the model. Nevertheless, the inclusion of the local loss allows for improved surface quality. Finally, with the introduction of multi-view regularization, the surface reconstruction performance is further enhanced, highlighting the superiority of our method in surface reconstruction.
• Figure 4: Comparison of visualization results. We presented rendered views of the test scenes, along with the corresponding depth maps and normal maps from the same viewpoint.
• Figure 5: Visualization of surface reconstruction. The figures showcase the results of training GigaGS on real aerial scenes, followed by rendering multiple RGB and depth maps, and ultimately obtaining the surface reconstruction results using TSDFusion zeng20173dmatch.