Towards Automating the Retrospective Generation of BIM Models: A Unified Framework for 3D Semantic Reconstruction of the Built Environment

TL;DR

SRBIM addresses the lack of a unified, scalable framework to convert 3D spatial and semantic information into BIM by introducing a three-component pipeline: PTv2-based semantic segmentation, MFS mesh reconstruction and refinement, and an IFC-based BIM reconstruction with partition-based color coding. The approach uses spatial features $F_g$ and color features $F_c$ in PTv2 to produce per-point semantic labels $\hat{L}_i$, partitions scenes into semantic segments, and reconstructs semantically enriched meshes via Poisson reconstruction, outlier filtering with threshold $\alpha=0.05$, and Laplacian smoothing, followed by mapping to IFC objects or IfcBuildingElementProxy and partition-coloring with average RGB $\overline{G}_{rgb}$. Evaluations across indoor, outdoor, landscape, and urban-scale datasets demonstrate SRBIM's ability to produce high-fidelity, colorized, semantically enriched BIM; results indicate robust performance with some over-aggregated surfaces. This pipeline promises automated BIM modeling suitable for real-world construction contexts and sets a new paradigm for automated BIM generation.

Abstract

The adoption of Building Information Modeling (BIM) is beneficial in construction projects. However, it faces challenges due to the lack of a unified and scalable framework for converting 3D model details into BIM. This paper introduces SRBIM, a unified semantic reconstruction architecture for BIM generation. Our approach's effectiveness is demonstrated through extensive qualitative and quantitative evaluations, establishing a new paradigm for automated BIM modeling.

Figures (4)

• Figure 1: Overview of the SRBIM pipeline. In the first step (a), the transformer model, PTv2, combines both spatial $F_{g}$ and color $F_{c}$ features to predict labels $\hat{L}i$. The model uses ground truth data $F_{gt}$ for supervised learning or as a substitute of $\hat{L}i$, with pseudo-labeling acting as an alternative to replacing $\hat{L}i$ when $F_{gt}$ is missing. This is followed by steps (b) and (c), where the initial point cloud scene is digitized into BIM.
• Figure 2: Qualitative results of MFS performance on example segments in the S3DIS hallway. Figs (a, c): semantic point cloud segments (Blue: Ceiling, Brown: Wall); Figs (b, d): Refined mesh segments produced by MFS.
• Figure 3: Intermediate results of MFS in the example segments from the S3DIS Hallway. Histograms (a, b) and heatmaps (c, d) compare the vertex density distributions of the mesh segment generated by Poisson Surface Reconstruction before and after post-filtering and smoothing.
• Figure 4: Reconstructed BIM model across datasets: All results are visualized in Blender blender_blenderorg_2024 with BlenderBIM blenderbim_blenderbim_2024 add-on. Color-enriched IFC objects are assembled to create a semantic-enriched BIM. The final BIM inherits all attributes and definitions of the IFC objects. IFC objects of distinct classes are instantiated, allowing them to be reused and modified. The circled examples of overextended surfaces indicate areas beyond the intended boundaries of the model structure.