jingwei-xu-00/eccv2026-cad-challenge-data

ECCV 2026 CAD Challenge Data

This challenge is part of the workshop The Path to Manufacturing: Evolving 3D Generation to Intelligent Computer-Aided Design.

Workshop homepage: https://3dgen-cad-workshop.github.io/

Challenge submission Space: https://huggingface.co/spaces/jingwei-xu-00/eccv2026-cad-challenge

This repository contains the public data package for the challenge. The evaluation Space accepts STEP predictions for the private evaluation split and updates the leaderboard after background evaluation.

Task At A Glance

Input: 3D render + TechDraw views.
Output: normalized STEP BRep.

Visual Examples

The examples below show training samples. Each row contains the primary 3D render, two hidden-line grayed (HLG) render views, and a README-only PNG preview rasterized from the corresponding techdraw/svg file.

Contents

• splits/train.csv and splits/train.jsonl: 8344 training samples with input paths and target_step_path.
• splits/test_public.csv and splits/test_public.jsonl: 927 public test samples with input paths only.
• train/render_3d/ and test_public/render_3d/: 3D render inputs, separated by split.
• train/techdraw/ and test_public/techdraw/: SVG/PDF/DXF TechDraw inputs.
• train/target_step/: normalized STEP targets for training samples only.
• examples/sample_submission.zip: format-only example submission containing one STEP prediction per public test sample.
• examples/min_eval/: minimum local evaluator smoke test using training targets.

Split Columns

Training rows contain:

sample_id, render_3d_path, render_3d_hlg_path, render_3d_hlg_translucent_faces_path, techdraw_svg_path, techdraw_pdf_path, techdraw_dxf_path, target_step_path

Public test rows contain:

sample_id, render_3d_path, render_3d_hlg_path, render_3d_hlg_translucent_faces_path, techdraw_svg_path, techdraw_pdf_path, techdraw_dxf_path

sample_id is split-local. Training IDs run from 000000 to 008343, and public-test IDs run from 000000 to 000926. Use (split, sample_id) as the unique key.

Download

hf download jingwei-xu-00/eccv2026-cad-challenge-data --repo-type dataset --local-dir eccv2026ws-cad-data

Minimum Local Evaluation Example

To smoke-test the bundled STEP evaluator locally, use examples/min_eval/. The helper copies a few training targets into matched gt/ and pred/ folders and then runs examples/min_eval/eval.py. Because pred == gt, the score should be near perfect if the evaluator environment is working.

cd eccv2026ws-cad-data
python examples/min_eval/run_min_eval.py --num-samples 2

The evaluator needs OCCT/PythonOCC and the same core Python packages used by the evaluation Space:

conda create -n eccv2026-cad-eval -c conda-forge \
  python=3.11 pythonocc-core numpy scipy trimesh tqdm
conda activate eccv2026-cad-eval

This smoke test has been run successfully with the eccv2026-cad-eval conda environment. A 2-sample run produced:

Valid Count: 2
Invalid Count: 0
Valid Ratio: 1.000000
Surface F1: 0.978260
Edge F1:    1.000000
Vertex F1:  1.000000
Topo F1:    0.983796
Summary:    0.990514

The exact surface/topology values may vary slightly because the metric samples geometry during evaluation.

Submission Format

Submit a zip file with one STEP file per public test sample:

submission.zip
  predictions/
    000000.step
    000001.step
    ...
    000926.step

To get an official test score, upload submission.zip to the evaluation Space. The public test answers are kept private there. For local experiments, create a validation split from the training set.

License And Attribution

See ACKNOWLEDGEMENTS_AND_LICENSES.md and metadata/license_attribution/ for the source and license information for all data used in this release. Please refer to the metadata file for the license of each individual asset.

Some assets may be subject to restrictions such as attribution, non-commercial use, or share-alike requirements.

Contact

For dataset bugs or challenge questions, contact davidxujw@gmail.com.