Source-linked AI summary
CRAB: Cuboid-based Representation for As-built BIM from Point Clouds
Haoran Xiong, Jing Wang, Shanshan Pan, Zihao Yan, Minglun Gong, Hui Huang
TL;DR
Point-cloud BIM reconstruction needs semantic, volumetric, editable components, but scans do not directly reveal key properties such as wall thickness and opening relationships. CRAB reconstructs architectural cuboids through proposal and relation-aware assembly, and reports stronger component reconstruction than floor-plan- and plane-based alternatives.
Problem
Raw scans provide noisy geometric measurements, while BIM requires editable semantic components, volumetric geometry, and relationships that are not directly observable.
Source: §1 · p. 2Method
CRAB represents architectural components as semantic cuboids and uses ProposalNet followed by AssemblyNet to refine geometry and relationships through relation-aware Flow Matching.
Source: §1 · p. 2, §4.3 · p. 5Results
CRAB reports more accurate wall-thickness estimation, more stable refinement, and more flexible BIM instantiation than floor-plan- or plane-based alternatives.
Source: §5.2 · p. 6, §6 · p. 8Takeaways & Limitations
Semantic cuboids provide editable architectural components and a direct interface for downstream BIM conversion.
Source: §6 · p. 8Entirely missing components may require coarse user initialization, and highly irregular or curved structures can be oversimplified by cuboids.
Source: §5.4 · p. 8
Abstract
from arXiv · showhide
As-built BIM reconstruction from point clouds demands semantic, volumetric, parametric, and editable building components, beyond 2D layouts or planar abstractions. Yet existing scan-to-BIM pipelines often rely on handcrafted rules, heuristic optimization, or manual correction, while learning-based methods fall short of recovering wall thickness, opening geometry, and component-level BIM structure. We introduce CRAB, a BIM-ready cuboid reconstruction framework that encodes walls, doors, windows, beams, and columns as compact architectural cuboids with IFC-compatible semantics and parameters. CRAB bridges point clouds and editable BIM by combining ProposalNet, which predicts coarse component cuboids, with AssemblyNet, a relation-aware Flow Matching model that refines geometry, predicts structural relationships, and enables automatic structural correction and human-guided editing. To address limited cuboid-level supervision, we extend Structured3D with component-wise cuboid annotations and construct cuboid-based procedural scenes. Experiments show CRAB surpasses floor plan and plane-based alternatives in component completeness, geometric accuracy, and generalization. For interior wall thickness estimation, CRAB achieves a mean error of 2.80 cm, with 79.50% of matched walls within 3 cm. The project page is available at https://vcc.tech/research/2026/CRAB.
1 Introduction
Point-cloud Scan-to-BIM methods often stop at layouts or planes, while BIM requires editable volumetric components with semantics and relationships. CRAB uses semantic cuboids and a proposal-and-assembly framework, supported by cuboid-level dataset construction.
1 Introduction
- Existing learning-based approaches often predict layouts or planes rather than volumetric component instances, limiting direct coupling of component geometry with BIM attributes.Raw scans also do not directly reveal wall thickness, wall-opening relationships, or editable instance identities.Source: §1 · p. 2
- CRAB represents walls, doors, windows, beams, and columns as semantic cuboids with parametric geometry and explicit topological relationships.Its centerline-based representation jointly encodes component semantics, wall adjacency, and opening-wall affiliation.Source: §1 · p. 2
- CRAB pairs ProposalNet's coarse semantic cuboids with AssemblyNet's relation-aware Flow Matching to refine geometry and predict component relationships.The refined cuboids retain editable parameters for user-guided modification and can be instantiated as IFC elements.Source: §1 · p. 2
- CrabSet addresses the shortage of cuboid-level supervision by adding component annotations to Structured3D and procedurally synthesizing architectural scenes.The dataset is constructed for training and evaluation of scan-to-BIM learning.Source: §1 · p. 2
- CRAB frames scan-based indoor BIM reconstruction as semantic cuboid assembly, explicitly modeling wall thickness, opening height, and editable volumetric geometry.This representation makes component geometry and semantics primary variables rather than deriving them from layouts or planes.Source: §1 · p. 2
2 Related Work
Prior indoor-structure methods produce floor plans, planes, component geometry, or editable cuboid programs, but these representations often leave BIM-ready volume or coordinated relationships implicit. CRAB motivates component-aware cuboids and local structural refinement.
2.1 Structured Indoor Modeling
- Floor-plan methods recover compact 2D layouts but do not directly specify wall thickness, vertical extent, opening height, or component volume.These quantities are required by CAD/BIM workflows.Source: §2.1 · p. 2
- Planar reconstructions require plane pairing and instance association to infer wall thickness, opening affiliation, and component identity, which can be ambiguous.CRAB instead reconstructs cuboid assemblies with explicit centerline, thickness, and height.Source: §2.1 · p. 2
2.2 BIM-Ready Cuboid Reconstruction
- Scan-to-BIM systems can recover component geometry and IFC-compatible outputs, but many rely on rules or workflows sensitive to clutter, occlusion, and incomplete scans.They also often omit explicit relations needed for coordinated refinement.Source: §2.2 · p. 3
- Existing cuboid and primitive methods mainly target object decomposition, shape generation, or furnished-scene synthesis, not architectural reconstruction from incomplete scene-level scans.CRAB uses centerline-based semantic cuboids and explicit relations for editable component reconstruction.Source: §2.2 · p. 3
2.3 Local Structural Refinement
- Feedback-driven editing and BIM agents motivate local correction, while CRAB's cuboid representation and AssemblyNet support user-guided refinement and global structural correction.The related approaches include code-render-inspect loops and local correction of missed or misplaced 3D layout elements.Source: §2.3 · p. 3
3 CrabSet Construction
CrabSet supplies point clouds, component-wise cuboid annotations, and pairwise architectural relations for BIM reconstruction. It combines Structured3D-derived residential scenes with procedurally synthesized architectural configurations.
3 CrabSet Construction
- Existing indoor datasets mainly provide floor plans, room layouts, semantic labels, or planar annotations, while explicit volumetric building components remain limited in scale or diversity.Table 1 compares availability of floor plans, structured 3D planes, and volumetric component-level annotations.Source: §3 · p. 3, Table 1 · p. 3
- CrabSet samples pair input point clouds with volumetric cuboid annotations and pairwise relations among architectural components.Cuboids record category, instance identity, position, orientation, and dimensions.Source: §3 · p. 3
- CrabSet combines Structured3D-derived residential annotations with synthesized scenes that vary boundaries, scales, and corridor-centered layouts.The synthesized examples cover non-uniform wall thicknesses, non-Manhattan contours, service shafts, and diverse circulation patterns.Source: Figure 2 · p. 3, §3 · p. 3
- The Structured3D-derived subset converts walls, doors, and windows into cuboids using planar faces, normals, semantic labels, and outer contours.It anchors CrabSet to a public residential benchmark.Source: §3 · p. 3
- Procedural subsets add large spaces with service shafts, columns, and multi-layer windows, as well as corridor layouts rarely represented in residential datasets.The scale-varied scenes target dense non-residential configurations; corridor-centered scenes include straight, bent, notched, H-shaped, T-shaped, and ring-like layouts.Source: §3 · p. 3
4 Methodology
CRAB formulates as-built reconstruction as semantic cuboids with explicit structural relationships, then proposes and assembles these components from point clouds. Its pipeline refines cuboids, predicts relationships, and converts the result into editable BIM components.
4.1 Problem Formulation
- CRAB represents walls, doors, windows, beams, and columns as parameterized semantic cuboids with explicit attachment and adjacency relationships.Cuboid geometry encodes a bottom central axis, thickness, and height, while the semantic category identifies the component type.Source: §4.1 · p. 4
4.2 ProposalNet
- ProposalNet learns to generate initial cuboid proposals from point clouds using a point-cloud encoder and an autoregressive cuboid decoder.The point-cloud features are adapted into the decoder token space before sequential cuboid generation.Source: §4.2 · p. 4
4.3 AssemblyNet
- AssemblyNet uses cuboid and point-cloud features to iteratively refine editable geometry while preserving fixed components, and predicts pairwise structural relationships.Its attention modules model within-cuboid and across-cuboid information and associate cuboid tokens with observed point-cloud geometry; predicted relations capture door-wall affiliation and wall-wall adjacency without hand-crafted inference rules.Source: §4.3 · p. 5
4.4 Reconstruction Pipeline
- The reconstruction pipeline proposes coarse cuboids, then lets users optionally add, delete, move, or anchor them before AssemblyNet refines geometry and relationships.Completely missing components require coarse user initialization; AssemblyNet then aligns and integrates them.Source: §4.4 · p. 5
- Refined cuboids and relations are converted into editable IFC components, with wall junctions resolved along wall length and door/window affiliations used for host-wall placement.Closed cycles of at least four wall segments are assembled into rooms, and wall structures can be instantiated as corner-joint, cross-joint, or wall-integrated types.Source: §4.4 · p. 5
5 Results and Evaluation
On Structured3D, CRAB improves component matching and volumetric geometry over floor-plan and plane-based alternatives, while also delivering competitive layout and plane quality. The evaluation further examines real-scene generalization, refinement, and parameterization; it notes limits in missing-component recovery, irregular geometry, and sparse component evaluation.
5.1 Experimental Setup
- The evaluation uses 3,000 Structured3D training scenes and 3,000 procedurally generated CrabSet scenes, with Structured3D validation and test splits and BIMNet and ISPRS for real-scene generalization.Baselines include volumetric, floor-plan, and plane-based methods; Ours* uses only S3D-derived cuboids while Ours uses CrabSet.Source: §5.1 · pp. 5–6
5.2 Evaluation
- 2.80 cm wall-thickness MAE and 79.50% T-Acc@3 show CRAB’s stronger volumetric reconstruction than alternatives, whose strongest non-cuboid T-Acc@3 is 41.62%.CRAB also improves window Match-F1 from 77.90% to 88.97% over SpatialLM and reports door and window height errors of 2.86 cm and 9.23 cm.Source: §5.2 · p. 6, Table 2 · p. 6
- CRAB achieves higher Corner F1 and Angle F1 than CAGE, comparable wall-plane performance to SpatialLM, and improves window F1 from 87.5% to 91.4% at IoU2D@0.5.CRAB’s gains in volumetric reconstruction do not come at the expense of layout or plane quality.Source: §5.2 · p. 6, Table 3 · p. 7
- Cuboid-based AssemblyNet improves interior-wall T-Acc@3 from 79.50% to 81.27%, whereas plane-based refinement can degrade its initial reconstruction.In simulated wall perturbations, cuboid-based refinement reduces endpoint errors to 3.60, 5.18, and 5.66 cm, compared with 30.29, 13.07, and 33.21 cm for plane-based refinement.Source: §5.2 · p. 7, Table 4 · p. 7, Figure 5 · p. 7
- On BIMNet, CRAB’s Match-F1 is 72.5% for walls, 68.8% for doors, and 48.5% for windows, versus SpatialLM’s 68.6%, 49.6%, and 37.4%.The reported trend also holds on residential-like BIMNet and non-residential ISPRS scenes.Source: §5.2 · p. 7
5.3 Ablation Study
- Centerline-based cuboids raise interior-wall T-Acc@3 from 71.74% to 79.50% versus oriented boxes, while oriented boxes perform slightly better on doors and comparably on windows.The centerline parameterization separates wall extent from thickness, a distinction suited to long, thin walls.Source: §5.3 · p. 8, Table 5 · p. 8
5.4 Limitation Analysis
- CRAB can correct continuous geometric errors, local misalignments, and redundant components, but entirely missing components may require coarse user initialization.Highly irregular or curved structures can be oversimplified by cuboids.Source: §5.4 · p. 8
- Sparse beams, columns, stairs, and service shafts in real-world test scenes limit comprehensive evaluation; robust reconstruction requires richer component-level supervision and broader BIM benchmarks.Source: §5.4 · p. 8
6 Conclusion and Future Work
CRAB represents architectural components as editable semantic cuboids with explicit geometric parameters, supporting downstream BIM conversion. Experiments report stronger volumetric reconstruction than floor-plan- or plane-based alternatives, while future work targets broader geometry and application scope.
6 Conclusion and Future Work
- Walls, doors, windows, beams, and columns are represented as semantic cuboids with centerline, thickness, height, and opening parameters for direct editing and downstream BIM conversion.Source: §6 · p. 8
- CRAB yields more accurate wall-thickness estimation, more stable refinement, and more flexible BIM instantiation than floor-plan- or plane-based alternatives, while remaining competitive in layout and plane quality.These findings are reported for synthetic and real-world scans.Source: §6 · p. 8
- Future work targets non-Manhattan geometry and richer primitives, broader BIM categories and structural relations, and interactive or agent-assisted workflows.The paper also proposes semantic plane cues as auxiliary evidence for reconstruction under real-scene domain shifts.Source: §6 · p. 8