3D vision · BIM automation

Scan-to-BIM: point cloud to verified IFC

A human-in-the-loop pipeline that turns a laser scan into an IFC model of slabs, walls and columns. Every element must be backed by real point support and a reviewer checks it before sign-off.

Context: AGECS (R&D) · contractor site scan and public benchmarks
Period: Jun 2026 – present
My role: R&D engineer: pipeline design, geometry classification, IFC generation, verification

Real contractor scan, plan-view point density: thin lattice members read as temporary shoring, so the pipeline keeps them out of the IFC.

~264 M
Points in the site scan: 7.7 GB PLY · single storey; 2 · 9 · 21
Slabs · walls · columns in IFC: re-verified with IfcOpenShell; 18.5 min
End-to-end regeneration: phased notebook, 0 blocking errors; 0.92
Column IoU (synthetic): PointNet++ · walls 0.89, floor 0.87

Chapter 01

The engineering challenge

Real scans are enormous: the site scan here was about 264 million points (7.7 GB). They also contain scaffolding, shoring and clutter that automatic pipelines mistake for structure.

An open-source baseline (Cloud2BIM) produced walls with the wrong orientation and forced-rectangle columns, which a reviewer rejected in ReCap.

Chapter 02

Input data

Site laser scan

~264 M points, 7.7 GB PLY, single storey (anonymised)

Public benchmark

Kladno station point cloud

Synthetic interior

207k labelled points for semantic segmentation tests

3D scatter of a segmented room point cloud — PointNet++ semantic segmentation output (synthetic interior).

Chapter 03

The AI & engineering workflow

Stage 1 of 5

Ingest

E57 / PLY → subsample (memory-mapped, chunked)

What I did

Adapted the open-source Cloud2BIM pipeline and kept its slab detection and IFC writer, while replacing wall and column classification with my own geometry classifier.
Fixed a ~90° wall-orientation bug (the `minAreaRect` angle convention) and checked it against independent PCA on 9 of 9 walls.
Built a multi-blob connected-component test that rejects temporary steel props and shoring without hard-coding to the dataset (23 → 21 columns, with all confirmed elements kept).
Wired the whole flow into a phased notebook with approval gates. It regenerates the IFC end-to-end in ~18.5 minutes, and I re-verified the output from a fresh IfcOpenShell process.
Benchmarked PointNet++ semantic segmentation on a synthetic building interior and ran the pipeline on the public Kladno station dataset.

PythonOpen3DNumPy / SciPyOpenCVIfcOpenShellPointNet++Autodesk ReCapRevit / Dynamo

Real scan · real IFC

The Kladno station scan and my final IFC model.

120,000 points sampled from the 250.5 M-point benchmark scan, coloured by the IFC element they support, then the model revealed storey by storey.

01 Scan

Real laser scan of Kladno railway station: 250.5 M points (120,000 shown).

My Scan-to-BIM engineDensity analysis + my geometry classifiers (PCA, connected components) → IFCFed: 250.5 M scan points

Real scan (public Kladno station benchmark) and my final IFC output, shown in their shared coordinates.

Drag to rotate · hover the model to see its supporting points

Input → output

Drag to compare.

Raw scanClass-labelled (ground truth)

Synthetic building interior: the unlabelled scan against its class-labelled version (walls, floor, ceiling, doors, columns).

Chapter 04

Technical result

Produced an interim IFC with 2 slabs, 9 walls and 21 columns. Each element traces to raw point support, and a fresh re-run reproduced the model exactly.
Removed temporary shoring and a lattice fragment automatically, keeping all 12 previously confirmed permanent elements.
Documented reusable methods: visual-first column-shape verification, and a temporary-works detection pipeline for future scans.

PointNet++ per-class IoU on a synthetic interior

Mean over the five classes present: 0.823.

The notebook's reported mIoU of 0.686 includes an empty background class scored as 0. Doors are most often confused with walls (16%).

View data table

Category	IoU
Column	0.922
Wall	0.888
Floor	0.874
Ceiling	0.873
Door	0.560

Source: AGECS PointNet++ evaluation (synthetic data)

Limitations, stated plainly

The site IFC is labelled interim. Openings (doors and windows) and one disputed element remain under review.

Outputs & artefacts

Five small renders isolating wall, floor, ceiling, door and column points — Fig. 01Per-class isolation used for QA.

Plan of a building footprint with labelled wall and column candidates — Fig. 02Contractor site scan: classified walls and columns over the slab footprint (coordinates removed).

Plot of slab candidate classification on a railway station scan — Fig. 03Public Kladno station benchmark: slab-candidate classification.

Chapter 05

Practical impact & relevance

Scan-to-BIM is costly manual modelling. This work automates the repetitive parts and keeps an engineer in control of what counts as structure.
Shows 3D geometry skills (PCA, contouring, connected components) alongside BIM data standards (IFC).

EvidenceWhere every figure on this page comes from

Each metric is taken from a primary record, not from a CV. Hover a metric to see its source. The underlying files are available for review at interview.

AGECS Scan-to-BIM progress log
AGECS Scan-to-BIM project record
AGECS PointNet++ evaluation (synthetic data)
AGECS Scan-to-BIM benchmark runs (Kladno)

The contractor and site are anonymised. Survey coordinates are cropped from all figures.

Next case study

Crack detection to digital twin: concrete water tank