The use of Gaussian splatting for 3D digitization for rock mass characterization and visualization

Abstract

This study evaluates the potential of 3D Gaussian Splatting (3DGS) as a novel 3D digitization method for rock mass characterization. The main focus was on the relative accuracy of 3DGS, its efficiency, and visualization capabilities compared to the traditional photogrammetric methods and manual field methods. Gaussian Splatting generally enables the creation of real-time, photorealistic 3D reconstructions through a point-based rendering technique that blends both geometry and texture data. This research investigated optimal data acquisition and processing workflows, the influence of device capabilities (mobile versus desktop), and the applicability of 3DGS under challenging underground conditions.

Using datasets obtained from the Underground Research Laboratory of Aalto University (URLA), models of 3D Gaussian splatting were made with Jawset Postshot software and geometrically compared to mesh-based photogrammetry and point cloud-based photogrammetry. Rock Quality Designation (RQD) measurements were extracted digitally with the three methods and compared with the manual/physical scanline measurements to assess geometric accuracy. Statistical analyses, including correlation coefficients and Root Mean Square Error (RMSE), were used to quantify accuracy among the different techniques.

Results indicate that 3D Gaussian Splatting achieved a higher correlation with the physical measurements compared to the mesh-based photogrammetric method and a subsequent lower RMSE to the mesh-based photogrammetric methods using the point cloud-based photogrammetry as reference. These results, therefore, showed the superior geometric fidelity of 3DGS to the mesh-based photogrammetric methods. Visual quality assessment by human participants also showed that Gaussian Splatting produced sharper, more complete, and colour-accurate representations. Although specular reflections and low-light noise presented challenges, 3DGS still outperformed traditional mesh-based photogrammetry in measurement consistency. These findings highlight 3D Gaussian Splatting as a promising tool for rapid, accurate, and interactive rock mass documentation and analysis in underground engineering applications.