@article{memari2026case,
  title = {A Case Study of Thermal Texture Transfer to Drone-Derived 3D Models with Neuralangelo and Instant-NGP: Using Photogrammetry, Neural Networks, and Data Fusion to Derive Thermal Twins},
  author = {Ammar MEMARI and Linus HENKEL and Jerome AGATER},
  year = 2026,
  url = {https://ibimapublishing.com/articles/CIBIMA/2026/224549/},
  journal = {Communications of the IBIMA},
  volume = 2026,
  pages = 14,
  doi = doi.org/10.5171/2026.224549,
  abstract = {Infrastructure diagnostics, energy efficiency analysis, and remote inspection can benefit from access to thermal 3D models of buildings and other installations. However, photogrammetrically reconstructing 3D models from thermal image data alone usually produces insufficient results due to a lack of texture, uniform regions, overexposure, and inadequate resolution. Our approach decomposes the problem into two parts: reconstruction of the 3D model and association of the thermal information with the 3D model. For reconstruction, we investigated the use of RGB imagery to photogrammetrically reconstruct the 3D model using COLMAP for traditional Structure-from-Motion (SfM), as well as the use of emerging neural reconstruction methods such as Neuralangelo. For subsequent thermal information transfer, we propose two fusion strategies: (1) Manual registration of image-pairs using an affine transform derived from user-selected correspondences to map thermal values to mesh vertex attributes; (2) Calibration and correction of lens distortion (radial/tangential, principal-point shift) and sensor misalignment to consistently project thermal imagery onto the reconstructed model. We use a real-world case study of a university service building at Jade Hochschule (Wilhelmshaven, Germany), with imagery captured by a DJI Mavic 3T drone, to validate the methodology, assessing the trade-offs in reconstruction fidelity, processing time, and thermal data accuracy. While our results show that vertex-based fusion is feasible, labor-intensive manual registration and information loss due to cropping and vertex sparsity limit accuracy and usability. Requiring only one-time parameter estimation, calibration and texture-based thermal projection yields more realistic, higher-fidelity thermal overlays, while enabling largely automated processing after the initial estimation. As a fast alternative for ad hoc thermal scene inspection, we evaluated Instant-NGP and 3D Gaussian splatting. While both can visualize thermal appearance, they suffer from reduced quality compared to our RGB-based versions and exhibit artifacts. Overall, our results indicate that an RGB-first reconstruction pipeline with camera-calibrated thermal texturing is the most viable path to usable thermal digital twins from drone data, while neural reconstructions may capture complex surfaces but remain computationally costly. Our modular, reproducible framework for thermal 3D digital twins is available as open source.},
  keywords = {Photogrammetry, Neuralangelo, Gaussian Splatting, thermal 3D reconstruction},
  note = Article ID: 224549
}