Generating IFC Models from 3D Gaussian Splatting (3DGS)

Semantic Enrichment of 3D Gaussian Splatting for IFC Model Generation

Bachelor Thesis Defense

Richard Würstlin

TH Köln – University of Applied Sciences

January 2026

Prof. Dr.-Ing. Niels Bartels | Dr. Yang Zou (University of Auckland)

Presentation Overview

1. Motivation & Problem Statement

2. Research Objectives

3. Theoretical Background

4. Developed Workflow

5. Tool Implementation

6. Results & Evaluation

7. Discussion & Limitations

8. Conclusion & Future Work

Motivation & Problem Statement

BIM requires semantic information beyond pure geometry

3D Gaussian Splatting offers fast, photorealistic capture

3DGS produces purely geometric models lacking IFC-required data

SEMANTIC GAP

How can 3DGS-derived data be enriched to generate structured IFC models?

Research Objectives

Bridge the gap between photorealistic 3DGS reconstructions and semantically rich IFC building models

Investigate how alphanumeric attributes can be integrated into 3DGS-derived geometries

Generate structured, semantically classified IFC models

Focus on shell construction elements: walls, floors, windows, doors

Develop and evaluate a workflow with supporting Python tools

Validate on real construction site dataset

3D Gaussian Splatting (3DGS)

Novel view synthesis method combining explicit scene representation with real-time rendering

Initializes from sparse point clouds via Structure-from-Motion (SfM)

Optimizes anisotropic 3D Gaussian primitives through differentiable rendering

Real-time rendering at 100+ fps at 1080p resolution

Training time under 1 hour vs. hours/days for NeRF

Fast acquisition + Photorealistic quality + Low equipment cost

Industry Foundation Classes (IFC)

Open standard for BIM data exchange (ISO 16739)

Layer 4: Domain Layer – Architecture, Structural Engineering, Facility Management

Layer 3: Interoperability Layer – Shared Elements: Floors, Walls, Doors, Windows

Layer 2: Core Layer – Product, Process, Relationship, Actor

Layer 1: Resource Layer – Geometry, Material, Properties, Quantities

IFC requires: Geometric information + Semantic classification + Alphanumeric attributes + Spatial relationships

Developed Workflow Overview

1. Reality Capture (3DGS)

2. Alignment & Noise Cleaning

3. Manual Segmentation

4. Patching Occlusion Holes

5. Voxel-Grid Downsampling

6. 3D Mesh Construction

7. PLY Export

8. IFC Generation

Input: 2.6M Points

Output: 15 IFC Elements (230 vertices)

Developed Python Tools

CC_cloud_patch.py

Fills holes in point clouds using convex hull + grid interpolation

CC_3D_mesh.py

Converts planar point clouds to extruded polygon meshes using PCA + bounding box

CC_mesh_export.py

Batch exports all mesh elements to PLY format preserving naming conventions

mesh2ifc.py

Converts PLY meshes to IFC with auto-classification, hierarchy generation & opening relationships

Tech Stack: Python | NumPy | ifcOpenShell | CloudCompare PythonRuntime | Tkinter GUI

3D Mesh Generation Tool

1. Extract points from selected point clouds into NumPy array

2. Fit plane using PCA to determine orientation

3. Calculate thickness based on point distribution along normal

4. Compute 2D contour (bounding rectangle or convex hull)

5. Create prism mesh by extruding the contour

Key Innovation: Recognizes walls as flat rectangular surfaces and reduces them to minimal geometry instead of thousands of triangles

IFC Export & Classification Tool

Auto-classification from filename keywords (wall, floor, window, door)

Complete IFC spatial hierarchy generation

Opening relationships via IfcRelVoidsElement + IfcRelFillElement

Geometry via IfcTriangulatedFaceSet

IfcProject → IfcSite → IfcBuilding → IfcBuildingStorey → IfcElement

Supported: IfcWall | IfcSlab | IfcWindow | IfcDoor

Results: Data Reduction Pipeline

Final IFC Model: 15 elements | 230 vertices | 400 triangles

Reduction Ratio: ~12,000:1 in point count

Results: Generated IFC Model

4 Walls (IfcWall)

2 Floors (IfcSlab)

7 Windows (IfcWindow)

2 Doors (IfcDoor)

Validated in BIMcollab Zoom: Complete spatial hierarchy + Correct opening relationships + Element classification

Estimated LOD: 200 (approximate dimensions and shapes with correct classification)

Mesh Comparison: Processing Stages

After Downsampling

400 Meshes | 230 Vertices

After Patching

324 Meshes | 192 Vertices

After Segmentation

388 Meshes | 224 Vertices

Key Finding: Visual differences are minimal - geometric outcome is primarily governed by boundary definition rather than raw point density

Limitations & Challenges

Data Loss in Format Conversion

Spherical harmonics, spatial extent, opacity discarded (~90% information reduction)

Manual Processing Requirements

~1 hour for single room with 15 elements; segmentation = 80% of time

Geometric Simplification

Assumes planar elements with uniform thickness; no curved/complex geometry

Single Room Validation

Tested on one construction site room; generalization requires further validation

Semantic Properties

IFC models contain geometry + classification only; no material/thermal properties

Future Research Directions

Automated Preprocessing

Specialized filtering for 3DGS-derived point clouds (floaters, edge artifacts)

Deep Learning Segmentation

Transfer learning from PointNet++/Swin3D; construction-specific training datasets

Mesh-Based Processing

Direct mesh workflows for surface definition, edge manipulation, topology

Semantic Property Inference

AI-based material estimation; database integration for property sets

Validation Extension

Multi-story buildings, complex geometries, different construction stages

Conclusion

Feasibility Demonstrated: 3DGS data can be transformed into semantically structured IFC models

Workflow Validated: Manual approach achieves 100% semantic correctness for shell construction

Tools Developed: Four Python scripts for patching, mesh generation, export, and IFC conversion

Data Reduction: 2.6M points → 15 IFC elements (230 vertices) = 12,000:1 ratio

Contribution: Methodological foundation for extending Scan-to-BIM to 3DGS-derived data

Outlook: Manual approach prioritizes reliability; automation requires domain-specific deep learning

Thank You

Questions & Discussion

Richard Würstlin

richard.wuerstlin@smail.th-koeln.de

Semantic Enrichment of 3D Gaussian Splatting for IFC Model Generation

Python scripts and digital appendix available on Sciebo platform