Catalog / Computer Vision / VIS-040

Scan-to-BIM & digital site survey (as-built models and floor plans from point clouds & phone scans)

Turn laser scans, photogrammetry or smartphone LiDAR captures into semantic as-built BIM/IFC models, floor plans and quantities for renovation, retrofit, quoting and facility management.

Parts needed to run it yourselfDocument readingVision

Existing buildings rarely have reliable drawings. Surveyors capture them with laser scanners or phones, then modellers spend days tracing walls, openings and pipes by hand. 3D semantic segmentation recognises building elements in the point cloud and outputs parametric BIM/IFC models, floor plans and measurement quantities, so that renovation, energy retrofit, trade quotes and facility management start from an accurate as-built model at a fraction of the manual modelling effort.

Problem & value

  • Today: the building stock in the EU is old and poorly documented; planners, energy consultants and trades re-measure on site (German: Aufmaß) with tape or laser distance meters, or commission a scan and a manual modelling service. Manual point-cloud tracing in Revit or ArchiCAD is slow and expensive, and dominates scan-to-BIM cost.
  • AI value: automatic classification of points into element classes, wall/slab/opening fitting, room segmentation and quantity take-off; for trades, a floor plan and surface areas within minutes of a phone scan, directly usable in a quote.
  • Policy driver: the EU Renovation Wave and the recast Energy Performance of Buildings Directive (EU) 2024/1275 require renovation passports, energy certificates and zero-emission building standards, which all need reliable geometry of existing buildings [1].
  • Benefit figures: vendors report large time savings versus manual modelling, but independent, comparable figures are scarce; measure your own baseline (hours per 1,000 m² modelled).

Variants

  • Scan-to-BIM for planners and surveyors: terrestrial or mobile SLAM scanner data to a discipline model (architecture, structure, MEP) at a defined level of detail (LOD) and level of accuracy (LOA, USIBD).
  • Smartphone room scanning and Aufmaß: phone/tablet LiDAR or camera captures produce floor plans, room dimensions and wall/floor areas for painters, floor layers, kitchen and window installers, insurance and estate agents.
  • Private renovators: consumer apps for measuring a flat before furniture or renovation planning.
  • Facility management / digital twin: scan-derived asset and room registers (CAFM/IWMS) with equipment classification.
  • Industrial plants: scan-to-CAD of piping and steel for brownfield revamps (piping fitting is a specialised sub-case; see DOC-013 for 2D P&IDs).

How it works

  1. capture (laser scanner / SLAM / phone LiDAR / photogrammetry)
  2. registration & cleaning (alignment, georeferencing, noise and dynamic-object removal)
  3. semantic + instance segmentation of points (wall, slab, door, window, column, pipe, duct, furniture)
  4. geometric fitting (planes, extrusions, openings, MEP runs; room/storey segmentation)
  5. parametric reconstruction (IFC / Revit families, floor plans, quantities)
  6. human QA against the cloud (deviation heat-map, manual correction)
  7. export (IFC, RVT, DWG/DXF, CAFM) and re-training from corrections
  • 3D segmentation: point-based or sparse-voxel networks (PointNet++, KPConv, Point Transformer V3, Sonata self-supervised pre-training) classify points; instance heads separate objects.
  • Vector reconstruction: rule-based and learned fitting converts labelled points into parametric walls with thickness, openings and storeys; occlusions (furniture, ceilings) are the main difficulty. Floor-plan methods from phone scans use room-layout estimation.
  • VLM role: vision-language models label unclear element types in images, read room signs and equipment tags from panorama frames, and check results against drawings.
  • Human-in-the-loop: a modeller reviews the deviation heat-map (model vs. cloud), fixes misclassified elements and signs off the model for its intended use. Survey-grade deliverables stay under a qualified surveyor's responsibility.

Data preparation

  • Data needed: laser-scan point clouds, photogrammetry image sets and phone LiDAR scans from site; optionally existing 2D plans, PDFs and IFC models from the client, the building archive or the DMS; element catalogues and your LOD/LOA specification.
  • Typical gaps: old plans only as paper or scanned PDFs, inconsistent file formats (E57, LAS, RCP, proprietary), scans without registration or georeferencing, furniture and people occluding walls, huge files scattered across laptops and file shares, and no agreed level of detail.
  • Preparation steps: (1) define LOD/LOA and element classes before scanning; (2) standardise a capture protocol (overlap, scanner positions, targets, lighting, time outside working hours); (3) digitise old plans (PREP-006) and convert formats (PREP-007); (4) register, clean and down-sample clouds, remove people and personal items; (5) organise projects with consistent naming and metadata on a central share (PREP-009, PREP-010); (6) keep modeller corrections as training data and a small reference set of verified rooms (PREP-016, PREP-018).
  • Tools: CloudCompare, PDAL, vendor registration software, IfcOpenShell, file-transfer to a NAS or EU cloud storage.
  • Share of effort: roughly 30-40 percent; capture quality, registration and storage dominate before any AI step.

Implementation paths

Which path for whom

  • SMB: A - mature scan-to-BIM tools and phone/tablet capture apps (several EU vendors) exist at project-based prices and the deliverables (floor plans, IFC) are standard.
  • Enterprise: A - commodity survey work is best bought; for portfolio-wide FM or integration with own CAFM/quoting, add Path B for the glue.
  • Developer: C - Pointcept, PDAL, CloudCompare and IfcOpenShell give a vendor-neutral pipeline and can be extended for special element classes.
  • Private: A - phone LiDAR apps (RoomPlan-based) produce usable floor plans for renovation planning without any infrastructure.
  • Phone/tablet capture: magicplan (Sensopia, France - EU), Apple RoomPlan-based apps (e.g. Polycam, Canvas, Matterport capture), PlanRadar for site documentation, Matterport for walkthrough twins with floor plans.
  • Scan-to-BIM automation: PointFuse (UK), Aurivus (Germany - EU), NavVis (Germany - EU; VLX mobile mapper and IVION platform), Faro Sphere XG / Scene, Leica Cyclone and Autodesk ReCap / Revit point-cloud tools, Trimble Business Center / Trimble Scan Essentials, Bentley iTwin Capture (ContextCapture), Cupix, Matterport scan services.
  • Check: E57/LAS/RCP import, IFC export quality, accuracy documentation (LOA), EU data residency, on-prem or tenant option and per-project pricing.
  • Choose A when: projects are recurring, deliverables are standard (architecture, basic MEP) and no in-house ML team exists.

Path B — Build on cloud APIs

  • Reality capture and 3D services: Autodesk Platform Services (Model Derivative, Reality Capture APIs), Trimble Connect/APIs, Bentley iTwin APIs, Esri ArcGIS reality tools and point-cloud classification (cloud-based; EU regions).
  • Glue with LLMs/VLMs: frontier VLMs (Claude, Gemini, GPT, Mistral Pixtral class, EU provider) on panorama stills for room naming, equipment labelling and report drafting; not for geometry.
  • Open IFC tooling: IfcOpenShell and the Speckle platform (open source, EU hostable) for writing IFC and syncing with Revit/Rhino.
  • Choose B when: you integrate scan outputs into your own quoting, FM or renovation workflow and want a vendor-neutral geometry layer.

Path C — Self-host (open source)

  • Point-cloud tools: CloudCompare, PDAL, Open3D, Potree for viewing, COLMAP / OpenMVS / Nerfstudio for photogrammetry, 3D Gaussian Splatting variants for visual capture.
  • Segmentation models: Pointcept (Point Transformer V3, Sonata, OA-CNNs), MinkowskiEngine / spconv backbones, Mask3D / OneFormer3D for instances. Benchmarks: ScanNet, ScanNet++, S3DIS, Structured3D, plus building-specific datasets (e.g. ARCH, BIMNet, Scan2BIM-type sets).
  • Output: IfcOpenShell for IFC authoring, FreeCAD BIM / Blender-BIM (BlenderBIM) for review.
  • Hardware: HW2 handles inference on room-scale clouds (tens of millions of points) after voxel down-sampling; HW3 (48 GB) for whole-building clouds (hundreds of millions of points).
  • Ops effort: high; maintain pipelines for registration, segmentation, fitting and QA.

Path D — Custom training

Worth it when you hold many labelled scans in a recurring building type (schools, offices, housing blocks, plants) and need element classes or LOD that off-the-shelf tools miss, for example MEP sub-types or legacy construction details. Needs hundreds of annotated rooms/floors; the corrections your modellers already make are the natural annotation source. Fine-tune a pre-trained backbone (Sonata/PTv3) instead of training from scratch.

State of the art (2026)

  • Models: transformer backbones with large-scale self-supervised pre-training (Point Transformer V3, Sonata) lead indoor benchmarks (ScanNet, S3DIS) and transfer reasonably to new buildings; open-vocabulary 3D segmentation (CLIP/VLM features lifted into 3D) allows querying "fire extinguisher" or "radiator" without retraining but is less accurate than supervised models.
  • Phone capture: Apple RoomPlan (LiDAR-equipped iPhones/iPads) returns parametric rooms with walls, doors, windows and furniture and exports USDZ; accuracy suffices for quotes and layouts, not for survey-grade as-built documentation [2].
  • Mobile mapping: SLAM handheld/backpack scanners (NavVis VLX, Leica, Faro, Stonex) cut capture time; AI-driven registration and automatic segmentation are now standard in vendor suites [3].
  • Beyond points: photogrammetry and Gaussian splatting give photo-real models, but extracting clean parametric BIM from them is still immature.
  • Limits and failure modes: occluded and hidden elements (above ceilings, inside walls) are not observable; glass, mirrors and shiny surfaces cause scan artefacts; thin or curved structures, heritage buildings and dense MEP clutter fail often; semantic class does not imply material or structural properties. Models need manual QA, and tolerance (LOA) must be agreed contractually.

Shared resources & combinations

  • Same stack (3d-vision + segmentation): VIS-026 (measurement, volume estimation), VIS-003 (drone infrastructure inspection with photogrammetry), VIS-019 (drone inventory), VIS-021 (aerial imagery) share photogrammetry, reconstruction and GPU infrastructure.
  • Upstream/downstream: ROBO-009 compares site progress against an existing BIM (this use case produces the as-built reference); DOC-013 reads 2D drawings as additional priors; PRED-048 uses the geometry for energy-performance and renovation roadmaps.
  • Data pipeline: the IFC model feeds quantity take-off, cost estimation, CAFM and digital twin platforms; the same labelled scan corpus supports Path D for multiple projects.

Effort, cost & KPIs

  • Path A: days to weeks; pay per project/area, per seat or per scan; plus capture hardware (phone LiDAR is cheap; SLAM/terrestrial scanners are a significant investment or a service).
  • Path B: 2-4 months for an integrated pipeline (capture → IFC → quoting/FM).
  • Path C: 4-9 months incl. GPU infrastructure and QA tooling; ongoing maintenance.
  • Cost drivers: scan capture time, registration, LOD/LOA required, manual QA effort.
  • KPIs: modelling hours per 1,000 m², element-level precision/recall, deviation model vs. cloud (mean/95th percentile in mm), share of elements accepted without correction, quote/survey turnaround time, rework rate on site.

People & change

  • Involve early: management (which project types), the project or survey lead, modellers and site staff who capture and review, IT (storage, large files, GPU or licences), the clients whose buildings are scanned, and the DPO when occupied rooms are captured. Define who signs off the model. See GUIDE-001, GUIDE-009 and GUIDE-010.
  • Works council: no in the typical case, because the system measures buildings, not staff. It becomes relevant if scans are made in occupied workplaces of your own company (employees visible) or if per-modeller productivity is evaluated: then §87(1) no. 6 and §90(1) no. 3 BetrVG apply and the council should be informed (GUIDE-004).
  • Data protection: scans of homes and offices capture people, documents and screens; capture when empty, blur, conclude an AVV with cloud vendors; a DPIA is rarely needed but check with the DPO (GUIDE-005).
  • AI literacy: support modellers and site teams with training on accuracy limits, occlusion errors and deviation heat-maps (Art. 4 AI Act; GUIDE-006).
  • Pilot and concerns: pilot on one or two recurring building types. Typical concern is that automation threatens modelling jobs or the surveyor's liability: show that modellers move from tracing to reviewing, the point cloud stays the legal source and a qualified person approves each model (GUIDE-008).

Risks, regulation & compliance

  • EU AI Act: minimal risk for building geometry; no Annex III category. Exception: if models feed safety components of critical infrastructure management, assess Annex III; AI literacy (Art. 4) applies in all cases.
  • GDPR: scans of occupied homes and offices can capture personal data (people, documents on desks, photos, screens); capture outside working hours, remove people and personal items, blur imagery and keep contracts and data processing agreements with vendors; prefer EU hosting.
  • Security and confidentiality: models of sensitive facilities (critical infrastructure, government buildings) need access control and, for NIS2 operators, supplier security review.
  • Liability: an AI-generated model is only as reliable as its stated accuracy; define LOD/LOA (USIBD, ISO 19650 information requirements), keep the point cloud as the legal source and have the model approved by a qualified person. In Germany, official cadastral or HOAI-relevant surveys remain with licensed surveyors.
  • Standards: IFC (ISO 16739), ISO 19650 information management, buildingSMART.
  • Mitigations: deviation heat-maps, spot checks with manual measurement, versioned audit trail.

Getting started

  1. Define the deliverable: LOD/LOA, formats (IFC, DWG), area per month, and who signs off.
  2. Pilot capture on 2-3 representative buildings with both a phone and a professional scanner.
  3. Test 2-3 Path A tools on the same data; compare element precision/recall and hours saved.
  4. Set up QA: deviation heat-map review and a tolerance table; log corrections.
  5. Connect output to quoting, CAFM or energy-assessment workflows; consider Path D after ~50 projects.

Sources

  1. European Commission - Energy Performance of Buildings Directive - https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficient-buildings/energy-performance-buildings-directive_en
  2. Apple Developer - RoomPlan - https://developer.apple.com/documentation/roomplan
  3. NavVis - Mobile mapping and reality capture - https://www.navvis.com/
  4. Pointcept - Point Transformer V3 / Sonata - https://github.com/Pointcept/Pointcept
  5. IfcOpenShell - https://ifcopenshell.org/
  6. buildingSMART - IFC standard - https://www.buildingsmart.org/standards/bsi-standards/industry-foundation-classes/
  7. magicplan - https://www.magicplan.app/
  8. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj