High-Quality Spatial Reconstruction and Orthoimage Generation Using Efficient 2D Gaussian Splatting

TL;DR

This work tackles the challenge of producing high-quality True Digital Orthophoto Maps without relying on Digital Surface Models or explicit occlusion detection. It introduces a 2D Gaussian Splatting (2DGS) framework that renders orthophotos and generates depth maps using an orthographic projection, aided by a divide-and-conquer training strategy to handle large-scale scenes efficiently. By leveraging planar Gaussian primitives and a batch rasterization rendering pipeline, the approach achieves sharp edges, faithful occlusion relationships, and accurate thin structures, with depth information enhancing spatial understanding for downstream tasks. The method demonstrates competitive performance against commercial software on UAV datasets, offering practical benefits for urban planning and environmental monitoring, albeit with longer training times that motivate future efficiency improvements.

Abstract

Highly accurate geometric precision and dense image features characterize True Digital Orthophoto Maps (TDOMs), which are in great demand for applications such as urban planning, infrastructure management, and environmental monitoring.Traditional TDOM generation methods need sophisticated processes, such as Digital Surface Models (DSM) and occlusion detection, which are computationally expensive and prone to errors.This work presents an alternative technique rooted in 2D Gaussian Splatting (2DGS), free of explicit DSM and occlusion detection. With depth map generation, spatial information for every pixel within the TDOM is retrieved and can reconstruct the scene with high precision. Divide-and-conquer strategy achieves excellent GS training and rendering with high-resolution TDOMs at a lower resource cost, which preserves higher quality of rendering on complex terrain and thin structure without a decrease in efficiency. Experimental results demonstrate the efficiency of large-scale scene reconstruction and high-precision terrain modeling. This approach provides accurate spatial data, which assists users in better planning and decision-making based on maps.

Figures (7)

• Figure 1: Illustration of our pipeline. The input consists of sparse point clouds and images with poses. After progressive data partitioning, training is conducted in parallel on different GPUs. Eventually, the trained Gaussians are projected onto an image plane using an orthographic projection method, with the complete TDOM rendered through batch rasterization. Simultaneously, the corresponding depth images are generated.
• Figure 2: Explanation of the differences between 3DGS and 2DGS. For 3DGS, the surfaces observed from different views are distinct and almost do not represent the actual surfaces, while the same plane is observed from any view with 2DGS.
• Figure 3: Illustration of two camera models. In the projection transformation, we need to move the cone frustum or square frustum to the coordinate origin and scale it to the range of [-1,1].
• Figure 4: Illustration of the degenerate solutions of 2DGS. We project the Gaussian position onto the image plane and create a standard 2D Gaussian centered at that point. By comparing the values at the intersection points of the current rays with the two Gaussians, we determine whether the 2D Gaussian has suffered from a degeneracy issue in the given view.
• Figure 5: An overview of the TDOM generated from the NPU DroneMap dataset and the Agisort sample dataset by our proposed method.