Unsupervised structural grouping for enhancing 3D reconstruction from investigation videos of Fukushima Daiichi Nuclear Power Station

This paper describes a three-dimensional (3D) reconstruction refinement pipeline based on unsupervised structural grouping. 3D reconstruction from underwater remotely operated vehicle videos is essential for investigation activities at the Fukushima Daiichi Nuclear Power Station. However, reconstruction from camera images alone is often degraded by noise, turbidity, floating particles, and reflections, resulting in artifacts and unstable results. Our method groups visually similar image regions without semantic annotations and computes a reconstruction contribution ratio for each group. An operator-in-the-loop graphical interface enables selective exclusion of unnecessary regions, and frame-wise masks are automatically generated and applied to an SfM/MVS-based reconstruction pipeline. Experiments using real Fukushima Daiichi investigation videos demonstrate that the proposed approach suppresses artifacts, improves reconstruction stability and interpretability, and introduces only limited computational overhead. These results show that the method is suitable for practical investigation scenarios requiring robust 3D reconstruction from degraded imagery.