Introduction
Generating True Digital Orthophoto Maps (TDOMs) is a cornerstone of modern aerial surveying. Unlike standard orthophotos, which often suffer from perspective distortion or "lean" in tall structures, TDOMs provide a perfectly nadir view where every pixel is accurately georeferenced. Traditional photogrammetric workflows rely on creating a Digital Surface Model (DSM) followed by complex, computationally expensive occlusion detection. These methods frequently struggle with blurred building boundaries and the loss of fine-grained structures like power lines.
This paper introduces a novel framework that bypasses explicit surface modeling by leveraging 2D Gaussian Splatting (2DGS) and differentiable rendering. By adopting a "divide-and-conquer" strategy, the proposed method achieves sharper reconstructions and superior handling of thin geometries compared to industry-standard commercial software.
Methodology
The core of the approach lies in transitioning from traditional 3D Gaussian ellipsoids to 2D elliptical disks. These 2D primitives align more naturally with the planar surfaces typical of terrain and building facades, reducing the perspective inconsistencies inherent in 3DGS for mapping applications.
System Workflow
The framework follows a multi-stage pipeline designed for scalability and precision, as shown in the workflow below.
- Progressive Data Partitioning: To manage large-scale drone datasets, the scene is divided into manageable cells based on camera positions.
- Parallel Training: Each cell is trained independently using 2DGS, allowing for efficient GPU utilization.
- Seamless Merging: The trained cells are integrated into a unified global model.
- Orthophoto Map Rendering: The final TDOM is generated using a specialized projection model.
Orthographic Projection
A key technical contribution is the modification of the standard 2DGS projection. While typical splatting assumes a pinhole camera model, this framework implements an orthographic camera model to render the top-down map directly.
This shift ensures that the rendered output is free from perspective distortion, a requirement for "True" orthophotos. Additionally, by calculating ray termination depths during splatting, the system produces high-precision depth maps alongside the visual imagery.
Results and Insights
The method was evaluated using the NPU DroneMap and Agisoft sample datasets, comparing performance against commercial tools like Metashape and Pix4Dmapper.
Qualitative and Quantitative Performance
The 2DGS approach demonstrated significant qualitative advantages, particularly in preserving "thin" geometry. While traditional mesh-based methods often "melt" structures like power tower supports into the ground, the splatting approach maintains these disconnected features with high fidelity. Quantitatively, the method achieved competitive geometric accuracy, with absolute distance errors typically under 1.0 meter when validated against Ground Control Points (GCPs).
Implicit Occlusion Handling
A major advantage of this differentiable rendering approach is that it handles occlusions implicitly. Because Gaussian primitives are depth-sorted and blended during training, the system naturally learns visibility without requiring separate, error-prone algorithms. This results in significantly sharper building edges and the elimination of "ghosting" artifacts common in traditional TDOMs.
Conclusion and Future Work
By leveraging the surface-centric nature of 2D Gaussian Splatting, this framework provides a robust alternative to traditional photogrammetry for high-quality spatial reconstruction. While the current implementation offers superior visual quality and geometric detail, the primary challenge remains training efficiency. Future efforts will focus on optimizing the computational pipeline to reduce the multi-hour GPU training time required for large-scale scenes, moving toward a more viable solution for rapid-turnaround industrial applications.
Summary updated