Mind the Domain Gap: Benchmarking Structure-from-Motion Methods on Paired Industrial Physical-Digital Twin Datasets to Enable Radiance Field Rendering

As industry transitions toward fully digitalized factories, mirroring physical environments in digital-twins has become essential for robotics, monitoring, and process automation, a capability that fundamentally relies on 6D camera pose estimation and 3D scene reconstruction. We propose an evaluation methodology to benchmark these tasks in paired physical and digital environments. Real-world acquisitions are performed on a diverse set of industrial objects – including a mobile robot, a robotic arm, a bike, and a bench drill – captured with a stereo camera, with accurate Ground-Truth (GT) poses provided by a Motion Capture (MoCap) system. To complement these acquisitions, we generate Digital-Twin (DT) counterparts by replaying the same camera trajectories in a similar Unity-based environment using the LOOPER tool. The resulting paired Physical-Twin (PT) and DT sequences share the same camera poses, enabling controlled analysis of the gap between real and simulated domains. We apply the methodology to evaluate several Structure-from-Motion (SfM) pipelines, assessing their pose accuracy. The resulting benchmark highlights the influence of image realism, object shape, and domain shift on pose estimation performance, providing a practical testbed for industrial vision research and supporting future work in 6D camera pose estimation, digital-twins, and 3D scene reconstruction.