Simulating virtual UAV surveys for benchmarking image-based 3D reconstruction methods: a comprehensive pipeline

The data and code in this repository can be used to simulate various scenarios for evaluating the 3D reconstruction quality of different image-based methods. To simulate a wide range of applications, two scenes with distinct characteristics were digitized: A high-alpine rock glacier and a forested area. In this context, digitization implies that a real-world scene was surveyed, e.g., using light detection and ranging (LiDAR) and photogrammetric techniques, and subsequently imported as a 3D mesh into the software Blender, where virtual image flights with varying acquisition parameters can be simulated.
Additionally, this repository contains Python scripts that enable fully automated planning of virtual flight missions, the processing of established photogrammetric methods (Structure from Motion (SfM) and Multi-View Stereo (MVS)) and novel radiance field methods (Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS)), as well as their conversion into point clouds, and the 3D evaluation of the results against the reference mesh. This enables analysis of 3D accuracy, completeness, and inlier ratio for each respective method.

This repository is part of a publication (Schulte et al., 2026). Detailed information on the processing and results is provided in this publication.