Research published by Spatial Source on 22 September 2026, from the Digital Twin and Land Tenure Lab at UNSW Sydney, tested whether better AI segmentation actually produces a better building model. The team compared five deep learning models (Swin3D, Point Transformer V3, Point Transformer V1, PointNeXt and PointMetaBase) on a modified Stanford Large-Scale Indoor Spaces dataset, then reconstructed walls with RANSAC and doors and windows with DBSCAN plus bounding boxes, using open-source tools including Open3D and IfcOpenShell. The headline result: Swin3D scored highest on wall segmentation at 88.95% Intersection-over-Union and also produced the most accurate walls, averaging 7.3 cm deviation from the reference cloud. But PointNeXt scored lowest on wall segmentation at 82.25% and still landed joint second-best on reconstruction at 7.7 cm. The same pattern held for openings: doors reached a 96.9% detection rate with the best model, windows only 75%, and the ranking on segmentation did not match the ranking on geometry. Read the original article for the full method and figures.
What this means for scan to bim work in Singapore
If you commission as-built models here, the practical takeaway is that the number a vendor shows you is probably the wrong number.
Segmentation scores describe how well a model labels points. They say nothing about whether the wall in the delivered IFC sits in the right place, whether the door opening is the right width, or whether a window that reads as glass in the scan was modelled at all. Those are the things that matter when the model goes into a submission, a tender, or a facilities handover.
The Singapore context sharpens this. Most of our Scan-to-BIM work is on buildings that were not designed to be modelled: older commercial floors with subsequent fit-outs, industrial buildings with services run above a false ceiling, conservation shophouses where the wall plane is not a plane. In those buildings, the elements the research flags as hardest are exactly the elements that carry the most risk. A missed window in a facade model affects a periodic facade inspection programme. A door modelled at the wrong width affects an accessibility assessment. A wall plane that is 7 cm out affects a clash check against new M&E.
The study's wall deviation figures are also worth reading carefully. 7.3 cm and 7.7 cm are averages against a reference cloud, on an indoor dataset, with a specific registration and a specific reconstruction method. That is not a tolerance you can carry across to a different scanner, a different building, or a different registration approach. Any accuracy figure you are quoted should come with the conditions it holds under: equipment, range, registration method, and what it was measured against.
What changes in practice, and what does not
What changes: how you evaluate a Scan-to-BIM deliverable.
| Old habit | Better habit |
|---|---|
| Ask what segmentation accuracy the tool achieves | Ask for deviation figures against the registered cloud, by element type |
| Accept a model viewport screenshot | Ask for a cloud-to-model comparison, with a colour deviation map |
| Treat "automatic" as a quality claim | Ask which elements were modelled manually and why |
| Judge the model as one object | Judge walls, doors, windows and services separately |
What does not change:
Automatic extraction still does not remove the need for a registered, controlled point cloud. If registration is loose, no amount of segmentation quality will rescue the geometry. The reconstruction step in the research works because enough correctly located points survive to define the main plane. If your cloud is smeared, the plane it finds is the wrong plane.
It also does not change the fact that openings and occluded areas need human review. The research is explicit that windows are the hardest class: small, less common, frequently occluded, and often co-planar with the wall behind them. Glass that is transparent or reflective is a capture problem before it is a modelling problem. No classifier fixes a window the scanner barely saw.
And it does not change the deliverable you should be asking for. If you need a model for a BCA submission, a renovation tender or an asset register, the acceptance criteria are geometric and informational, not statistical. The model either matches the building within an agreed tolerance or it does not.
Our view
The research is a useful correction to a habit we see in procurement: buying a tool's benchmark rather than a deliverable.
Segmentation benchmarks are cheap to publish and easy to compare, which is why they dominate the conversation. Reconstruction quality is harder to score and depends on the building, so it gets less airtime. The UNSW result that a lower-scoring model can produce better walls is not a quirk. It reflects something anyone who has registered a real scan knows: the reconstruction step is tolerant of noise as long as the structure is intact, and intolerant of a missing plane no matter how clean the labels are.
Where we would push back on reading this as a green light for full automation: the study was run on indoor datasets, with a modified training set, and the cross-dataset test on 11 Matterport3D rooms showed performance dropping when the environments differed from the training data. Singapore buildings differ from Stanford indoor spaces in ways that matter. Concrete soffits with service penetrations, tiled wet areas, mirrored lift lobbies, curtain wall with coated glass, and the general density of M&E above a false ceiling all sit outside the easy cases. A model that performs well on a research dataset is not evidence it will perform well on your building.
The open question the paper raises but does not settle is what "good enough" means for reconstruction. If a workflow can produce a usable model from imperfect segmentation, the industry needs a way to state the tolerance that model was built to, and to test it. Right now most as-built specifications say something like "accurate to 20 mm" without saying at what range, with what equipment, or against what reference. That is the gap worth closing, and it is a specification problem more than a research problem.
Who should care: anyone writing a Scan-to-BIM scope, anyone accepting a model at handover, and anyone budgeting for a renovation where the model drives the design. If you are scanning a single room for a fit-out, this is mostly academic. If you are modelling a 20,000 sq m industrial building with the intent of running a digital twin off it, the quality of the reconstruction is the whole project.
One honest note on when not to hire us: if your actual need is a set of measured drawings for a small A&A job, a full Scan-to-BIM model may be more than the job requires. A registered point cloud plus targeted 2D sections is often faster and cheaper. We will say so if that is what your scope looks like.
What to do next
If you are scoping a Scan-to-BIM job, write the acceptance criteria before you write the scope of works. Specify the tolerance, the element types it applies to, and how it will be verified. Ask for a deviation map against the cloud, not a segmentation score.
- For the modelling side, see Scan-to-BIM.
- For the capture that feeds it, see 3D laser scanning and drone and aerial LiDAR where the scope includes roofs or facades.
- If the end use is an operational model rather than a construction one, see digital twin capture.
- If you already own a scanner and want to produce your own cloud, our equipment rental line covers Lixel K2, K1 and L2 Pro, PortalCam, Matterport Pro3 and Leica BLK360 G2. You run the capture and produce your own outputs.
- To talk through a scope, request a quote.
