Digital Twin from Point Clouds in Existing Buildings

A point cloud is not automatically a digital twin. It is raw material — millions to billions of points with no semantic meaning. The route from there to a usable BIM as-built model is craft, experience and good software. Here is the workflow we use — and the pitfalls that trip up so many first-time projects.

What a point cloud actually is

A point cloud is a list of 3D coordinates (X, Y, Z), optionally with colour and reflectance values. A typical E57 dataset from a Matterport Pro3 scan of a 1,000 m² office contains roughly 50–150 million points. Those points "know" nothing about what they are — there is no distinction between wall, window, door, furniture or pot plant.

Step 1: Cleaning and decimation

Before any further processing, the raw point cloud is cleaned: people, moving objects and scan artefacts are removed. It is then decimated — a controlled, uniform reduction that preserves edges and detail while stripping out redundancy. A hundred million points often comes down to 20–30 million, a range any normal CAD workstation can handle smoothly.

Step 2: Classification

Classification tools (in our pipeline usually CloudCompare, Recap Pro or FME) assign points to semantic categories — floor, wall, ceiling, window, door, furniture. This is where operator experience matters: the algorithms can do a lot, but complex geometries (suspended ceilings with ventilation diffusers, for example) still need manual correction.

Step 3: Modelling

The classified point cloud is then turned into the actual model inside BIM software (Revit, ArchiCAD or OpenBIM tools). Every wall, every window, every door is rebuilt as a proper BIM object — with properties such as material, build-up and layer thicknesses. This is genuine manual work, not a one-click conversion.

Typical effort: a straightforward 100 m² office area takes 6–10 hours of modelling time. A complex period-building floor with vaulted ceilings and irregular angles can easily cost three to four times that.

Step 4: Validation

The finished BIM model is validated against the point cloud: overlay, deviation analysis, plausibility check. Acceptable tolerance is typically ±2–3 cm. Anyone who needs better than that is asking for a different survey class than our standard.

Level of Detail — what you actually need

For most existing-building BIM projects, LOD 200–300 is the right level of detail. LOD 400 is really only worth it on specialist projects with high-precision pre-fab planning.

Commercials: what BIM as-built modelling really costs

On our side, modelling effort is priced by the hour. A complete 800 m² office floor at LOD 300 typically lands at €6,000–12,000 in modelling costs — on top of the capture itself (Matterpak is already included there). That is markedly more expensive than the scan itself — which is exactly why we only recommend BIM when the downstream workflow genuinely needs it.

Common pitfalls

1. Unclear brief: "BIM model" without an LOD specification almost always leads to rework.
2. Wrong software choice: a model authored in Revit is not 1:1 transferable into ArchiCAD.
3. Stale point cloud: if months pass between scan and modelling and the building has changed in the meantime, the model becomes unusable.
4. Over-specified LOD: asking for LOD 400 because it "sounds professional" costs three to four times as much without delivering proportionate value.

Conclusion

A digital twin built from a point cloud is craft work — it takes time and costs accordingly. It is worth it for the use cases where the workflow genuinely demands it: complex refurbishments, BIM-coordinated new-build extensions to existing structures, large industrial-plant projects. For most smaller schemes, a point cloud plus a 2D floor plan is the pragmatic middle ground.

We give you an honest opinion on whether your project actually needs as-built BIM — even when the answer is that it does not: Request consultation.