GeomCAD

Dataset for GeomCAD: A Benchmark for Precision Reconstruction of Generative CAD Operations from Point Clouds.

Layout

extrude_100w/       # training set (to be uploaded)
loft_100w/          # training set (to be uploaded)
revolve_100w/       # training set (to be uploaded)
sweep_100w/         # training set (to be uploaded)
test/
  extrude/          # 1500 test samples
    test_mesh_paths.json   # {"<id>": "<id>/<id>.stl"} relative to test/extrude/
    <id>/<id>.stl          # ground-truth mesh
    <id>/<id>.py           # CAD-Recode-style code that produces the mesh
  loft/             # 1500 test samples
  revolve/          # 1500 test samples
  sweep/            # 1500 test samples

split	operation	#samples	size
test	extrude	1500	61 MB
test	loft	1500	353 MB
test	revolve	1500	409 MB
test	sweep	1500	410 MB

Usage

from huggingface_hub import snapshot_download
local = snapshot_download("FanQY/GeomCAD", repo_type="dataset")

License

This dataset is distributed under CC-BY-NC-4.0, matching CAD-Recode v1.5's licensing for upstream compatibility. Generation and evaluation code are released separately under the MIT license.

@inproceedings{geomcad2026,
  title={GeomCAD: A Procedural Benchmark for Multi-Operation Point-Cloud-to-CAD Reconstruction},
  author={{Fan, Qinyuan and Yu, Wentao and Xiong, Jian and Sandfeld, Stefan and Gühmann, Clemens},},
  year={2026}
}