Tortho-Gaussian: Splatting True Digital Orthophoto Maps

In this paper, we propose a novel method for generating True Digital Orthophoto Maps (TDOMs) using 3D Gaussian Splatting (3DGS). Traditional methods for TDOM generation face challenges such as occlusion detection and distortion correction, which often result in inaccuracies and visual artifacts. Our approach leverages 3DGS to directly produce orthophotos by projecting Gaussian ellipsoid kernels onto 2D image planes, bypassing the need for complex occlusion detection procedures. Additionally, we introduce a divide-and-conquer strategy to efficiently manage large-scale scenes, optimizing both memory usage and rendering time. Complementing this approach, we implement a fully anisotropic Gaussian kernel that adapts to the varying characteristics of different regions, thereby enhancing accuracy and visual quality, particularly in the representation of reflective surfaces and slender structures. The proposed method significantly improves the accuracy of building edge reconstruction and eliminates distortions in weak texture regions. Experimental results demonstrate that our approach outperforms existing commercial software in both visual quality and computational efficiency, making it a promising solution for large-scale urban scene reconstruction.