AMADEUS – 3D Foot Reconstruction Pipeline

AMADEUS (AI‑based Morphological Analysis & Design Engine for Unique Shoes) reconstructs a personalized 3D foot model from smartphone video and produces a ready-to-print shoe last. It combines YOLO11n-seg + SAM for segmentation, COLMAP for camera pose estimation, 3D Gaussian Splatting (3DGS) for point cloud optimisation, and SuGaR for mesh generation【528303309685221†L80-L94】.

Pipeline

1. Pre‑processing: Segment the foot and scale marker from each frame and save masked images【717972134425363†screenshot】.
2. 3D Reconstruction: Use COLMAP to build a sparse point cloud【894071447112061†screenshot】.
3. Undistortion & Alignment: Remove lens distortion and align the sparse model【894071447112061†screenshot】.
4. 3DGS Training: Optimise a Gaussian point cloud to produce a dense representation【35837800387949†screenshot】.
5. Meshing & Healing: Convert the Gaussian field to a watertight mesh and remove noise【756719135168637†screenshot】.
6. Scaling: Convert to real-world dimensions using a checkerboard marker.
7. 3D Printing: Slice and print the mesh.
8. Troubleshooting: Fix segmentation resolution mismatches and ensure adequate data【756719135168637†screenshot】.

Installation and Usage

Docker

docker build -t amadeus .
docker run --gpus all -it --rm \
  -v $(pwd)/data:/app/data \
  -v $(pwd)/output:/app/output \
  amadeus
chmod +x run_pipeline.sh
xvfb-run -a ./run_pipeline.sh

Manual Setup

Clone submodules (gaussian-splatting and SuGaR), install dependencies from requirements.txt, install COLMAP, and run the commands in run_pipeline.sh manually.

Citation

If you use this project, please cite:

@report{amadeus2025,
  title = {AI‑based Morphological Analysis & Design Engine for Unique Shoes},
  author = {Kim Taeryang and Park Hyundong and Park Chanwoo and Bang Hojun},
  year = {2025},
  note = {P–Practical Project third semester (AI) team 3 report}
}

License

MIT License.