Yytsi/floorplan-to-3d-walls

Floorplan to 3D — segmentation model

A UNet decoder with a ResNet-34 encoder backbone (ImageNet-pretrained), trained on CubiCasa5K to segment architectural floor plans into four classes: floor / wall / door / window.

The full project (training code, polygon extractor, FastAPI inference server, 3D Three.js viewer) lives at https://github.com/Yytsi/floorplan-to-3d.

A hosted preview at floorplan-to-3d.pages.dev shows three pre-rendered example plans extruded into 3D — it's static, no backend. To run inference on your own SVGs, clone the repo and start the local FastAPI server (the viewer detects it via /healthz and reveals an upload button).

Files

• best.safetensors — model weights, ~98 MB. Loadable with safetensors.torch.load_file. The epoch and best_miou are stored in the safetensors metadata header.
• config.yaml — training config required by the loader (image size, letterbox, normalization, encoder name).

Quick start

mkdir -p weights
curl -L -o weights/best.safetensors https://huggingface.co/Yytsi/floorplan-to-3d-walls/resolve/main/best.safetensors
curl -L -o weights/config.yaml      https://huggingface.co/Yytsi/floorplan-to-3d-walls/resolve/main/config.yaml

Then follow the README on GitHub for the inference server + viewer.

Training

• Data: CubiCasa5K (high-quality + high-quality-architectural splits).
• Input: 512×512 RGB, aspect-preserving letterbox so non-square plans aren't stretched. ImageNet-mean fill in the padding.
• Encoder: ResNet-34 pretrained on ImageNet.
• Loss: cross-entropy.
• Best val mIoU: 0.983 at epoch 26.

Limitations

• CubiCasa5K is dominated by Nordic residential plans drawn in a specific CAD style. The model generalizes well within that distribution but may miss elements on hand-drawn sketches, commercial blueprints, or plans using very different line conventions.
• Only the four structural classes above are predicted. Furniture, fixtures, and text annotations are treated as floor.
• Input is rasterized from SVG at 512×512; very long thin walls in oversized plans can get sub-pixel and be missed.

License

MIT — see the GitHub repo.