A Reproducible Reconstruction and Evaluation Framework for Photorealistic Lunar Analog Scene Reconstruction Toward Digital Twins

This research presents a pipeline to quantify photorealistic scene reconstruction accuracy from monocular camera feeds against physical ground truth in lunar analog settings. The authors benchmark six variants of 3D Gaussian Splatting across three scenes, demonstrating that standard training metrics are insufficient predictors of cross-reconstruction photorealistic novel view synthesis fidelity. They introduce a Coverage Score heuristic based on COLMAP poses and point cloud structure to assess structural fidelity independently of splat training performance. The work establishes a standardized evaluation pipeline for lunar-analog Digital Twin development, with code available at the provided project page.