Skip to content

Multimodal and embodied data

Point Clouds Paired with BIM Models for Scan-to-BIM AI

Quick answer

A usable scan-to-BIM dataset pairs a registered point cloud of a real building with the as-built BIM model that was authored from that scan, plus enough correspondence to know which points belong to which wall, slab, duct or column. Buyers should specify the capture method, registration accuracy, model level of development, element-level linkage, building owner and client authorization, and how occupants and sensitive interiors were screened. Public benchmarks rarely cover these pairs at production scale, so most training-grade pairs come from firms that delivered scan-to-BIM work.

By SourceX Editorial · Updated

What a training-grade scan and BIM pair contains

A training-grade pair has three parts: the registered point cloud, the BIM model at a stated level of development, and an explicit mapping between them. The point cloud usually arrives as E57, the E57, the ASTM E2807 exchange format, which stores 3D points, per-point color and intensity, and the 2D images captured by the scanner in one vendor-neutral file. Native scanner project files (for example registered scan projects from the capture software) carry more metadata but tie you to one toolchain, so ask for both where available.

The model should be IFC, the buildingSMART schema for building objects, properties and relationships, now published as ISO 16739-1:2024 for IFC 4.3. A Revit or ArchiCAD native file is useful as a second copy, but IFC exports give you stable GlobalIds and classes such as IfcWall, IfcSlab, IfcDoor and IfcFlowSegment that become your label vocabulary. The third part, correspondence, is where most delivered projects fall short: the modeler fit geometry to the cloud but rarely recorded which points support which element.

How to get element-level correspondence without relabeling everything

The practical answer is to derive point labels from the model, then audit a sample by hand. Once the cloud and model share one coordinate system, each point can be assigned to the nearest IFC element within a distance threshold, with unassigned points kept as clutter or "not modeled." That produces per-point GlobalId and class labels for segmentation training and per-element point sets for scan-to-BIM fitting.

Failure modes are predictable. Models are often built at LOD 300 to a stated tolerance, so a wall drawn as a perfect plane will not match a bowed real wall, and furniture, MEP above ceilings and temporary works are frequently omitted. Ask whether the modeler logged deviations or clashes; some teams record those as BCF topics tied to element IDs, which is a useful quality signal and a source of "hard case" labels [1]. Also ask whether the model was later updated for design changes, because a model that drifted away from the scan date silently corrupts every derived label.

Capture metadata to require per scan

Every scan should carry its sensor, density and registration evidence, because model performance on new sites depends on matching those conditions. Terrestrial laser scanners, mobile and SLAM-based scanners, and photogrammetry produce very different noise, occlusion and density patterns, and a model trained on one transfers poorly to another. Registration reports (target or cloud-to-cloud residuals, georeferencing, control points) tell you whether a few centimeters of misalignment will bleed into your labels.

Illustrative example: invented to show structure; it does not describe an available dataset.

FieldExample valueWhy it matters
pair_idPR-0007Joins cloud, model and audit records
building_useHospital outpatient wing, 3 floorsDomain shift and sensitivity screening
capture_methodTerrestrial laser scanner, 41 setupsNoise and occlusion profile
point_formatE57, color and intensity, registeredTool-neutral loading
registration_reportCloud-to-cloud residual summary attachedBounds label error from misalignment
model_formatIFC 4.3 plus native authoring fileStable GlobalIds and classes
stated_lodLOD 300 architecture, LOD 200 MEPWhich classes are trustworthy
model_toleranceStated by modeler in scope documentDistance threshold for label transfer
correspondenceDerived per-point GlobalId, 2% hand-auditedSegmentation-ready labels
scan_date / model_dateSame month; no later design updatesAvoids drift between cloud and model
privacy_screeningFaces and screens removed from images; patient areas excludedBiometric and confidentiality risk
rights_chainOwner authorization letter; scanning firm and modeler consentLicense can be granted

Package the manifest with Croissant-style metadata so each record's files and fields are machine-readable for your loaders [4]. For the full request layout across modalities, use the multimodal dataset specification template.

Why public point-cloud benchmarks are not enough

Public indoor benchmarks are valuable for pretraining and comparison but do not cover the pairing problem. The Stanford 2D-3D-S release, which includes the widely used S3DIS scans, is labeled indoor data from a handful of mostly academic and office buildings. It has semantic point labels, not an authored BIM model with element identities, levels of development or MEP systems, and its terms of use are written for research, so confirm whether commercial training is permitted.

Production scan-to-BIM covers hospitals, plants, retail, data centers and historic buildings, with ceiling voids, plant rooms and exteriors that benchmarks seldom contain. If your model will be sold to AEC firms, you need private pairs across building types, and you also need a private held-out set so evaluation reflects real projects; see private evaluation sets for multimodal models.

Who can license the pair: owners, clients and modelers

Rights in a scan-to-BIM pair usually sit with several parties at once, so the license must trace each of them. The building owner controls access and often owns the deliverables by contract; the scanning or engineering firm that captured the data and the modeler who authored the BIM may retain rights in their work product unless the services agreement assigned them. Ask for the services agreement clause covering deliverable ownership and reuse, or a written authorization from the client, before treating a firm's project archive as licensable.

Copyright in the building design is a separate question. US copyright law (17 U.S.C. 120) limits an architect's control over pictures of a constructed building only where the building is located in or ordinarily visible from a public place, and that limit does not obviously reach interior scans or the BIM model, which may embody the design or the modeler's own work. Multi-party licensing patterns are covered in licensing multimodal records from several rightsholders.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Screening occupants and sensitive interiors

Reality capture records whatever was in the room, so screening is part of preparation, not an afterthought. Panoramic images embedded in E57 files and colorized points can show faces, monitors, whiteboards, patient names and documents on desks. Faces matter legally: Illinois BIPA treats a scan of face geometry as a biometric identifier and requires a retention schedule and written release [2], and Texas requires notice and consent before capturing face geometry for a commercial purpose [3].

Ask suppliers to state what was removed (embedded images, color, specific rooms), how faces and screens were masked, and how that was checked. Security-sensitive spaces such as server rooms, secure labs, correctional facilities or critical infrastructure may need to be excluded entirely or generalized. Combined image, audio and metadata screening is covered in de-identifying multimodal records.

Buyer checklist before you request data

Write the request around the model you are training, not around a building type. A precise request lets a supplier answer yes or no quickly and makes pricing comparable.

  • Target task: per-point semantic segmentation, instance segmentation, parametric element fitting, or full cloud-to-IFC generation.
  • Classes and LOD: which IFC classes must be modeled, and at what level of development and tolerance.
  • Capture mix: scanner types, photogrammetry or mobile capture, interior or exterior, occupied or empty.
  • Correspondence: derived labels acceptable, or hand-labeled subsets required, with an audit rate.
  • Building mix: uses, ages, sizes and regions; exclusions such as healthcare or secure facilities.
  • Rights: owner or client authorization, services-agreement ownership terms, and modeler consent.
  • Privacy: removal of faces, screens and documents; treatment of embedded imagery.
  • Delivery: E57 and IFC with native files where possible, manifest, and checksums.

Related engineering data types live on different pages: 2D drawings paired with 3D CAD models for mechanical drawing-to-model work, CAD and PCB design datasets for product design files, and construction project records for RFIs, submittals and schedules. Firms that hold this data are typically architecture firms and engineering consultancies. The multimodal data hub and the AI data guide index cover other paired modalities.

How SourceX handles scan-to-BIM requests

SourceX sources operational datasets from US companies on request; it does not hold scan-to-BIM data in stock, and a request does not guarantee a match. Buyers describe the data they need, and SourceX looks for US businesses that hold it, with every release approved by the supplying company. Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery, through private, access-controlled workflows after an executed agreement. You can describe your scan-to-BIM data requirements on the buyers page.

Request paired scan and BIM data for your model

If your team needs registered point clouds paired with as-built BIM models, describe the classes, LOD, capture mix and rights you need. SourceX assesses data and licensing permissions with potential suppliers, and nothing is contracted until a supplier agrees. Start at https://sourcex.si/buyers.

Sources

  1. buildingSMART International, "BIM Collaboration Format (BCF)". https://technical.buildingsmart.org/standards/bcf/
  2. Illinois General Assembly, "Biometric Information Privacy Act (740 ILCS 14/)". https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004
  3. Texas Legislature, "Texas Business and Commerce Code Section 503.001 - Capture or Use of Biometric Identifier". https://statutes.capitol.texas.gov/Docs/BC/htm/BC.503.htm
  4. MLCommons Croissant working group (arXiv), "Croissant: A Metadata Format for ML-Ready Datasets" (2024). https://arxiv.org/pdf/2403.19546

Tell us what your models need

Share scope, volume, language, format, timing and licensing requirements.

Request data