A Comprehensive Evaluation of the Spatial Accuracy of Building Gaussian Splatting

Abstract. 3D building models are powerful visual tools, typically generated with well-established image-matching or LiDAR methods. However, they do not capture the view-dependent colour characteristics possible with Gaussian splatting. Despite the visual potential of Gaussian splatting, there is limited knowledge on its spatial accuracy and influencing factors, particularly for buildings. To address this gap, a two-building dataset was collected with terrestrial laser scans, images, phone LiDAR, and target points, and the visual and spatial effects of numerous factors were analyzed. These factors included the source and quality of the input camera poses and point cloud, the number of images and training iterations, and the Gaussian splat method. Gaussian splats were trained from open source and commercial image-based reconstruction methods, COLMAP and Pix4D, and phone LiDAR reconstructions. Applying Gaussian splatting to these inputs had minimal impact on the target points and the overall structure of the buildings, but the positions of Gaussians deviated from the initial point cloud, particularly before 15,000 iterations, resulting in more floaters and lower spatial accuracy. Image-based reconstruction methods outperformed phone LiDAR methods on visual and spatial metrics. Cleaning COLMAP point clouds considerably decreased Gaussian floaters, while downsampling input point clouds increased the percentage of floaters and yielded similar visual results. 2D Gaussian splatting provided geometric constraints, removing some floaters, but sacrificed visual quality. Increasing the number of images to three loops around the buildings improved visual and spatial results. Overall, the spatial accuracy of building Gaussian splatting was heavily dependent on the factors studied.