Abstract
True Digital Orthophoto Maps (TDOMs) are critical for Geographic Information Systems (GIS), providing a nadir view with uniform scale by removing perspective distortions. However, traditional TDOM generation is hindered by complex occlusion detection and reliance on Digital Surface Models (DSMs). This paper introduces Tortho-Gaussian, a framework that rethinks mapmaking by adapting 3D Gaussian Splatting (3DGS) for orthographic projection. By replacing traditional occlusion detection with an orthogonal splatting technique, Tortho-Gaussian achieves high mapping accuracy and visual quality for large-scale urban environments while bypassing traditional bottlenecks.
Introduction
In GIS, a TDOM is the gold standard for aerial representation. Unlike standard aerial photography, every pixel in a TDOM has a uniform scale, effectively removing the "leaning" effect of buildings. Traditionally, this requires a complex pipeline involving DSM generation, rigorous occlusion detection, and image rectification. These steps often fail at building edges or with reflective surfaces, leading to "ghosting" artifacts. Tortho-Gaussian addresses these issues by leveraging point-based radiance fields and a novel orthographic rendering approach.
Methodology
The core of Tortho-Gaussian lies in three primary innovations that adapt the 3DGS framework for professional surveying.
Orthogonal Splatting and Occlusion Handling
Standard 3DGS relies on perspective projection. Tortho-Gaussian derives a new mathematical formulation for orthographic splatting. By projecting 3D Gaussian ellipsoids directly onto a 2D map plane along the Z-axis, the method handles occlusions inherently through depth-sorting. This eliminates the need for explicit DSMs or manual occlusion checks, which are common failure points in traditional photogrammetry.
Fully Anisotropic Gaussian Kernel (FAGK)
To capture fine geometric details like power lines and reflective facades, the authors introduced a "plug-and-play" Fully Anisotropic Gaussian Kernel. This kernel integrates transparency, scaling, and rotation parameters into spherical harmonic representations. This allows the model to better capture view-dependent effects and maintain sharp edges, preventing the blurring often seen in neural rendering of aerial imagery.
Divide-and-Conquer Scalability
To manage the high memory requirements of 3DGS for city-scale mapping, Tortho-Gaussian employs a divide-and-conquer strategy. The scene is partitioned into overlapping rectangular cells, each optimized independently. These cells are then seamlessly merged, allowing the system to scale to large areas without exceeding single-GPU VRAM limits.
Experimental Results
The method was evaluated on the NPU-DroneMap and WHU datasets, including the high-precision "Chengzi" sequence.
- Precision: Tortho-Gaussian achieved an average relative error of 0.155% compared to industry-standard software like Agisoft Metashape, confirming its viability for professional surveying.
- Visual Quality: Qualitative analysis showed significantly cleaner building edges and the total elimination of facade shifts (the "leaning building" effect).
- Efficiency: The scalability strategy proved essential. On the "Phantom-ieu" sequence, Tortho-Gaussian completed training in approximately 39 minutes using 10.9GB of VRAM, whereas standard 3DGS failed due to memory exhaustion.
Conclusion and Future Work
Tortho-Gaussian demonstrates that representing a scene as a field of 3D Gaussians allows for a more elegant solution to the occlusion problem in mapmaking. By bypassing the need for perfect 3D meshes, it creates visually and geometrically consistent maps. Future research will focus on integrating lighting models to remove shadows and incorporating semantic segmentation (e.g., using SAM) to distinguish between permanent structures and transient objects like vehicles.
Summary updated