CRAB: From point clouds to an editable BIM model, with a wall thickness estimated to be within 2.8 cm

Walls, doors, windows, beams and poles are reconstructed as IFC-compatible parametric parallelepipedes, which can be corrected by hand.

Producing the digital model "as built" (as built BIM) of an existing building from a laser survey remains largely manual: scan-to-BIM chains are based on handwritten rules or human corrections, and learning methods struggle to regain wall thickness, aperture geometry or component structure.

Propose, then assemble

CRAB represents every wall, door, window, beam and pole as a compact parallelepiped, with a semantics and parameters compatible with the IFC format. A first network, ProposalNet, offers coarse volumes; a second, AssemblyNet, based on flow matching, refines geometry, deduces structural relationships between elements and allows both automatic correction and editing guided by a human. In the absence of annotated data at this level, the authors enriched the Structured3D game and generated procedural scenes.

CRAB surpasses approaches based on floor plans or geometrical planes in completeness and accuracy. For the thickness of the interior partitions, the average error is 2.80 cm, and 79.5% of the matched walls are estimated to be within 3 cm.