IFCNet GCN Autoencoder
a graph autoencoder for 3D building elements
2 min readupdated
This model analyzes the 3D geometry of building elements from IFC (Industry Foundation Classes) files to detect anomalies and find similar parts. Taking an input point mesh, it reconstructs the geometry and compresses it into a latent vector (a compact digital representation of an object). Built on graph convolutions (GCN — a graph convolutional network that processes data as connected points and graphs), the architecture has roughly 1.7 million parameters. Created as a research prototype for architectural analysis, it is useful for BIM and CAD specialists.
Key features
- Graph-based processing: Analyzes shapes using kNN graphs (k-nearest neighbors graphs) built from surface points.
- Anomaly detection: Compares the input point cloud with the reconstructed output using Chamfer distance (a distance metric between point clouds).
- Latent representation: Generates a 32-dimensional vector for finding similar elements via cosine distance.
- ONNX format: Runs easily locally via onnxruntime without heavy dependencies.
What you can do
- Validate elements: Automatically check building components for deviations from standard IFCNet classes.
- Find similar parts: Search for matching architectural components using the cosine similarity of their latent vectors.
- Integrate into pipelines: Use the repository scripts to preprocess IFC files and run inference.
- Test via API: Send requests to the Gradio demo server for quick results.
Example
import json
import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
import ifc_graph
repo = "iaac-maai/hackathon-team2-ifcnet-gcn"
cfg = json.load(open(hf_hub_download(repo, "config.json")))
sess = ort.InferenceSession(hf_hub_download(repo, "model.onnx"))
points, edge_index, _ = ifc_graph.ifc_to_arrays("element.ifc")
recon, latent = sess.run(None, {"x": points, "edge_index": edge_index})
print("latent shape:", latent.shape)
Read more
Details• Architecture: GCN (3 layers) + MLP decoder, 1,745,184 parameters.
• Benchmarks: ROC AUC 0.609, Precision 0.261, Recall 0.245, P@5 0.331.
• Training: 30 epochs on CPU using the IFCNetCore dataset (7,930 files).
• Limitations: processes a single element per IFC file, and size is normalized (scale is ignored).










Comments