Weeks to Turn a Scan Into a Model

Renovation and adaptive reuse of existing buildings requires accurate as-built digital models, but the majority of existing building stock has no BIM model and no original drawings. 3D laser scanning captures geometry as dense point clouds with millimeter accuracy, but converting these into semantically rich BIM models — identifying walls, floors, columns, MEP systems, and their relationships — remains overwhelmingly manual. A skilled technician spends 2–4 weeks processing a single medium-sized building scan. Historic and older buildings include irregular shapes, non-standard dimensions, and hidden structural elements that defeat pattern-matching approaches.

Renovation accounts for 55–70% of construction activity in developed countries by value. Without accurate as-built models, renovation design relies on incomplete drawings and field measurements, leading to design clashes, rework, and cost overruns averaging 20–30%. The EU's Renovation Wave aims to double building renovation rates by 2030, but the modeling bottleneck limits how fast the existing stock can be digitized. Millions of buildings need renovation for energy performance and the BIM pipeline cannot keep pace.

Automated scan-to-BIM software (Autodesk ReCap, Trimble RealWorks, ClearEdge3D) can extract planar surfaces and regular geometric primitives but fails on curved surfaces, complex intersections, and MEP routing. Deep learning for point cloud semantic segmentation (PointNet, PointNet++, RandLA-Net) shows promise on benchmark datasets but degrades sharply on real buildings where training data is sparse and building typologies diverse. Multiple site visits are often required because single scans have occlusions — areas behind equipment, inside ceiling plenums, or obscured by temporary conditions. Resulting BIM models contain systematic errors (misaligned elements, missing connections, incorrect material assignments) that propagate into renovation design.

Large, diverse scan-to-BIM training datasets covering multiple building typologies (residential, commercial, industrial, historic) with paired point clouds and ground-truth BIM models. Algorithms that reason about occluded geometry — inferring hidden structural elements from visible evidence using building typology priors and construction-era conventions. Interactive AI-assisted workflows combining automated extraction with efficient human-in-the-loop correction, reducing the 2–4 week process to days.

A team could prototype an AI-assisted point cloud segmentation tool for one building typology (e.g., mid-century residential), training on scans collected from campus buildings with known floor plans. Alternatively, a team could develop an occlusion-inference algorithm that uses visible structural clues (exposed beam ends, wall thicknesses) to predict hidden framing configurations. Computer vision, BIM/CAD, and architecture skills apply.