3D Gaussian Splatting for Large-scale Surface Reconstruction from Aerial Images

TL;DR

A novel 3DGS-based method for large-scale surface reconstruction using aerial multi-view stereo (MVS) images, named Aerial Gaussian Splatting (AGS), which can match conventional aerial MVS methods on geometric accuracy in aerial large-scale surface reconstruction and beats state-of-the-art GS-based methods both on geometry and rendering quality.

Abstract

Recently, 3D Gaussian Splatting (3DGS) has demonstrated excellent ability in small-scale 3D surface reconstruction. However, extending 3DGS to large-scale scenes remains a significant challenge. To address this gap, we propose a novel 3DGS-based method for large-scale surface reconstruction using aerial multi-view stereo (MVS) images, named Aerial Gaussian Splatting (AGS). First, we introduce a data chunking method tailored for large-scale aerial images, making 3DGS feasible for surface reconstruction over extensive scenes. Second, we integrate the Ray-Gaussian Intersection method into 3DGS to obtain depth and normal information. Finally, we implement multi-view geometric consistency constraints to enhance the geometric consistency across different views. Our experiments on multiple datasets demonstrate, for the first time, the 3DGS-based method can match conventional aerial MVS methods on geometric accuracy in aerial large-scale surface reconstruction, and our method also beats state-of-the-art GS-based methods both on geometry and rendering quality.

Figures (12)

• Figure 1: The overview of AGS Framework. (1) The SfM sparse point clouds and views are divided into N data blocks. (2) The point clouds in each block are used to initialize the 3D Gaussians. (3) The ray-gaussian intersection technique is applied to obtain depth and normal vector information. (4) The depth map and normal map are utilized to compute the depth normal consistency constraints and multi-view geometric consistency constraints.
• Figure 2: Overview of Adaptive Aerial Scene Partitioning strategy. (a) The entire scene is divided into N regions based on camera positions. (b) The boundaries of each region are expanded. (c) Viewpoints (i.e., cameras) are selected and culled. (d) All points visible from the selected viewpoints within each data block are added to the block’s point cloud.
• Figure 3: Ray-Gaussian Intersection. By calculating the maximum Gaussian value along the ray, we can obtain accurate depth and normal vector information for the Gaussian.
• Figure 4: Multi-view geometric consistency constraints. The multi-view geometric consistency constraints are modeled as the error between projection and reprojection of depth map across multiple views.
• Figure 5: Surface reconstruction results of WHU-OMVS dataset.