Tortho-Gaussian: Splatting True Digital Orthophoto Maps

TL;DR

This work introduces Tortho-Gaussian, a method to generate True Digital Orthophoto Maps by orthogonally splatting optimized 3D Gaussian fields, eliminating the need for explicit DSM and occlusion detection. It combines a divide-and-conquer strategy for large-scale scenes with a Fully Anisotropic Gaussian Kernel that uses spherical harmonics to adapt appearance and geometry to different regions, improving reflections and slender structures. The approach yields higher fidelity building edges, continuous facades, and better performance in weak-texture areas, demonstrated across urban datasets with favorable efficiency and memory usage compared to vanilla 3DGS and commercial tools. The results indicate strong potential for scalable, high-quality TDOM production in digital twin and GIS workflows, with future directions including city-scale extension and integration of semantic and depth information.

Abstract

True Digital Orthophoto Maps (TDOMs) are essential products for digital twins and Geographic Information Systems (GIS). Traditionally, TDOM generation involves a complex set of traditional photogrammetric process, which may deteriorate due to various challenges, including inaccurate Digital Surface Model (DSM), degenerated occlusion detections, and visual artifacts in weak texture regions and reflective surfaces, etc. To address these challenges, we introduce TOrtho-Gaussian, a novel method inspired by 3D Gaussian Splatting (3DGS) that generates TDOMs through orthogonal splatting of optimized anisotropic Gaussian kernel. More specifically, we first simplify the orthophoto generation by orthographically splatting the Gaussian kernels onto 2D image planes, formulating a geometrically elegant solution that avoids the need for explicit DSM and occlusion detection. Second, to produce TDOM of large-scale area, a divide-and-conquer strategy is adopted to optimize memory usage and time efficiency of training and rendering for 3DGS. Lastly, we design a fully anisotropic Gaussian kernel that adapts to the varying characteristics of different regions, particularly improving the rendering quality of reflective surfaces and slender structures. Extensive experimental evaluations demonstrate that our method outperforms existing commercial software in several aspects, including the accuracy of building boundaries, the visual quality of low-texture regions and building facades. These results underscore the potential of our approach for large-scale urban scene reconstruction, offering a robust alternative for enhancing TDOM quality and scalability.

Figures (13)

• Figure 1: A simplified model of building. It includes roofs and ground marked with different colors. The blue and green cameras represent perspective projection, with their imaging containing the gray building facades. The red camera represents orthographic projection.
• Figure 2: Toy examples for generating TDOM. (a) The traditional photogrammetric TDOM generation from multi-view aerial images, incorporating an occlusion detection. Two images I and II are captured from two perceptive centers, with projection lines illustrating occlusion relationships from various viewpoints. In this solution, the DBM (Digital Building Model) and DEM (Digital Elevation Model) are required as input. These assist in occlusion detection, ensuring that only visible building surfaces are displayed, while eliminating façade shifts caused by viewpoint changes. (b) The proposed Tortho-Gaussian, generating a seamless, complete TDOM without the need for mosaicking. Scale-nonuniform and original images are used for 3DGS training. Then, an orthographic projection of the building cluster is performed along the z-axis, bypassing occlusion detection. The Gaussian ellipsoid field is rendered at a selected spatial resolution to produce a true orthophoto of the entire scene. This approach eliminates the need for post-processing steps such as image mosaicking and radiometric/color corrections, offering a streamlined and efficient solution.
• Figure 3: Workflow of the proposed Tortho-Gaussian. First, it begins by aligning the sparse point cloud to the x- and y- axes. The scene is then partitioned into smaller regions using a divide-and-conquer strategy. For each partition, the original 3D Gaussian Splatting (3DGS) framework is employed to train the corresponding Gaussian field with the enhanced Fully Anisotropic Gaussian Kernel (FAGK). The trained Gaussian fields are seamlessly merged into a unified field to represent the entire scene, eliminating the need for TDOM tiling and subsequent color and brightness balancing steps. Next, the camera position is selected, and 3D Gaussian ellipsoids are splatted orthogonally, followed by pixel rasterization to compute the color for each pixel on the TDOM of the complete scene.
• Figure 4: Divide-and-Conquer strategy. The dark rectangles are the initial divisions of the scene, the light-colored rectangles indicate the extended regions, and the orange and purple dots represent cameras selected by the local region and the external region, respectively.
• Figure 5: The complete TDOMs of the NPU-DroneMap and WHU dataset using our proposed Tortho-Gaussian.