Scan-to-BIM, the conversion of terrestrial laser scanning point clouds into as-built Building Information Models (BIMs), is the main route to geometrically reliable digital records of existing buildings. A controlled 2021 study by Esfahani, Rausch, Sharif, Chen, Haas and Adey quantified the dimensional accuracy and precision of BIMs produced by 25 modellers in five manual and semi-automated modelling scenarios, but on a single small university laboratory containing relatively simple objects. That study explicitly leaves open two linked questions: (1) whether modeller training also improves accuracy and precision for geometrically complex secondary objects (ducts, pipes), which it only conjectured; and (2) how a hybrid workflow combining manual and semi-automated modelling affects dimensional accuracy, precision, recall, and modelling time, including whether it recovers the recall deficit of purely semi-automated modelling. No study resolving these two questions is known in the literature to date, so the problem remains open.

Background

Building Information Modeling (BIM) is the practice of representing a building as a parametric digital model whose elements (walls, slabs, ducts, pipes) carry geometry and semantics. Accurate as-built models of existing buildings are needed to operate, maintain, and adapt facilities, but most existing buildings lack such models. Scan-to-BIM is the dominant approach for creating them and has three phases: (1) scanning, typically with a terrestrial laser scanner (TLS), an instrument that measures dense 3D coordinates of surfaces to produce a point cloud; (2) registration, the alignment of scans from multiple stations into one coordinate system; and (3) modelling, the conversion of the point cloud into parametric BIM objects. Modelling is the current bottleneck of the process and dominates its time cost. It is performed either manually - a modeller visually superimposes parametric objects onto the point cloud in software such as Autodesk Revit - or semi-automatically, where software such as EdgeWise (which uses iterative fitting algorithms like RANSAC, a robust model-fitting method) detects objects automatically and a human corrects the results. In this literature, accuracy is the mean absolute dimensional error of modelled element dimensions relative to ground-truth dimensions; precision is the standard deviation of that error across repeated trials; and recall is the fraction of required objects that are actually modelled, with manual modellers conventionally assumed to achieve 100% recall. Primary objects are the principal building surfaces (walls, floors, ceilings); secondary objects are geometrically more complex elements, here the duct-and-pipe system. In the source study, a point cloud of a small University of Waterloo laboratory was captured with a Faro Focus M70 scanner (ranging accuracy of plus or minus 3 mm); 25 modellers modelled the scene in five scenarios: manual modelling of primary objects before and after receiving training, manual modelling of secondary objects, and semi-automated modelling of primary and of secondary objects. Ground-truth dimensions came from laser distance meter measurements using the visible-surface convention, and the error distributions were fitted with normal distributions to estimate the probability that a modelled dimension falls within a given error range. The study found that training improved the average accuracy and precision of primary-object BIMs by 260% and 330% respectively; semi-automation improved trained-manual accuracy and precision by 20% and 1455% for primary objects but only 9% and 36% for secondary objects; and that different legitimate modelling assumptions (for example fitting a wall to its innermost versus outermost scanned points) produce very different dimensions from the same point cloud, so two adequately trained modellers can produce equally correct but dimensionally divergent BIMs. Because recall of the semi-automated software is lower than that of human modellers while its accuracy and precision are higher, and because training was tested only for primary objects, the authors identified an unresolved modelling-side agenda: quantify the untested factor combinations (training for secondary objects; hybrid manual-plus-semi-automated workflows) and work toward a unified system for standardizing and certifying scan-to-BIM outputs, which industry still lacks; the study's error statistics are confined to that single simple laboratory case.

Problem Statement

The problem is the unfinished quantitative evaluation of the modelling phase of Scan-to-BIM posed by Esfahani et al. (2021). Within their experimental framework - dimensional error against visible-surface ground-truth measurements, accuracy as mean absolute error, precision as the standard deviation of error, recall as the fraction of required objects modelled, and per-scenario modelling time - determine: (1) whether and by how much modeller training improves the accuracy and precision of modelling secondary objects (ducts, pipes), which the source study could only conjecture to behave like primary objects; (2) how a hybrid modelling workflow, in which manual and semi-automated software modelling are combined (for example semi-automated detection with manual completion or correction), affects dimensional accuracy, precision, recall, and modelling time relative to the pure manual scenarios (manual modelling of primary objects before and after receiving training, and manual modelling of secondary objects) and the pure semi-automated scenarios, including whether the hybrid recovers the recall deficit of purely semi-automated modelling while keeping its accuracy, precision, and time advantages. These two components share one scientific objective: to give practitioners and standards bodies evidence-based dimensional error bounds for scan-to-BIM outputs as a function of the modelling workflow, which is the prerequisite for the unified standardization and certification of scan-to-BIM outputs that the source authors call for. Component (3) - whether the reported error magnitudes and the probability that a modelled dimension falls within a given error range generalize beyond a single small laboratory containing relatively simple primary and secondary objects to larger, more complex, or more cluttered buildings - and the standardization and certification of scan-to-BIM outputs are out of scope for this problem and provide context only. The problem concerns a controlled empirical study, within the source's metric framework, that answers (1) and (2), reporting error statistics per object class and modelling scenario.

References

  1. Mansour Esnaashary Esfahani, Christopher Rausch, Mohammad Mahdi Sharif, Qian Chen, Carl Haas, Bryan T. Adey (2021). Quantitative investigation on the accuracy and precision of Scan-to-BIM under different modelling scenarios. Automation in Construction 126:103686. DOI 10.1016/j.autcon.2021.103686, https://doi.org/10.1016/j.autcon.2021.103686
  2. Yu-Chen Lee, Jong Won Ma, Fernanda Leite (2023). A parametric approach towards semi-automated 3D as-built modeling. Journal of Information Technology in Construction 28:806-825. DOI 10.36680/j.itcon.2023.041, https://doi.org/10.36680/j.itcon.2023.041
  3. Nuno Abreu, Andry Pinto, Anibal Matos, Miguel Pires (2023). Procedural Point Cloud Modelling in Scan-to-BIM and Scan-vs-BIM Applications: A Review. ISPRS International Journal of Geo-Information 12(7):260. DOI 10.3390/ijgi12070260, https://doi.org/10.3390/ijgi12070260
  4. Qian Wang, Jiajun Li, Xiaoying Tang, Xinchen Zhang (2022). How data quality affects model quality in scan-to-BIM: A case study of MEP scenes. Automation in Construction 144:104598. DOI 10.1016/j.autcon.2022.104598, https://doi.org/10.1016/j.autcon.2022.104598
  5. Jidnyasa Patil, Mohsen Kalantari (2025). Automatic Scan-to-BIM - The Impact of Semantic Segmentation Accuracy. Buildings 15(7):1126. DOI 10.3390/buildings15071126, https://doi.org/10.3390/buildings15071126
  6. Jidnyasa Patil, Arcot Sowmya, Mohsen Kalantari (2026). Automatic Scan-to-BIM: The Impact of Semantic Segmentation Accuracy on Opening Detection. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences XI-1-2026:437-445. DOI 10.5194/isprs-annals-xi-1-2026-437-2026, https://doi.org/10.5194/isprs-annals-xi-1-2026-437-2026
  7. Yudong Liu, Han Huang, Ge Gao, Ziyi Ke, Shengtao Li, Ming Gu (2025). Dataset and benchmark for as-built BIM reconstruction from real-world point cloud. Automation in Construction 173:106096. DOI 10.1016/j.autcon.2025.106096, https://doi.org/10.1016/j.autcon.2025.106096
  8. Mudan Wang, Zhiqi Hu, Hamidreza Alavi, Guangming Wang, Ioannis Brilakis (2025). Cebim: A Benchmark Dataset for Scan-to-BIM of Buildings from Point Cloud to IFC Models. SSRN preprint. DOI 10.2139/ssrn.5162663, https://doi.org/10.2139/ssrn.5162663
  9. M. Jarząbek-Rychard, H.-G. Maas (2023). Modeling of 3D geometry uncertainty in Scan-to-BIM automatic indoor reconstruction. Automation in Construction 154:105002. DOI 10.1016/j.autcon.2023.105002, https://doi.org/10.1016/j.autcon.2023.105002
  10. Rnin Salah, Nóra Géczy, Kitti Ajtayné Károlyfi (2026). Predictive hybrid scan-to-BIM method improves heritage building documentation completeness and accuracy. Scientific Reports 16:7622. DOI 10.1038/s41598-026-38200-8, https://doi.org/10.1038/s41598-026-38200-8
  11. Gustavo Rocha, Luís Mateus (2021). A Survey of Scan-to-BIM Practices in the AEC Industry - A Quantitative Analysis. ISPRS International Journal of Geo-Information 10(8):564. DOI 10.3390/ijgi10080564, https://doi.org/10.3390/ijgi10080564
  12. Tristan Kinnen, Jörg Blankenbach (2026). Automated geometry-based analysis for object-level Scan-vs-BIM validation of TBE components. Journal of Information Technology in Construction 31:75-105. DOI 10.36680/j.itcon.2026.004, https://doi.org/10.36680/j.itcon.2026.004
  13. Jidnyasa Patil, Mohsen Kalantari (2025). Automatic Scan-to-BIM: A Systematic Analysis of Challenges, Research Gaps, and Future Directions. SSRN preprint. DOI 10.2139/ssrn.5639196, https://doi.org/10.2139/ssrn.5639196