CADRec

Reconstructing a CAD Sequence Recursively with Localized Geometric Contexts

ACM TOG (SIGGRAPH Asia 2026)

Haoxuan Song, Bingchen Yang, Jun Xiao, Haiyong Jiang

Project Page · Paper (PDF) · Video

CADRec reconstructs editable CAD sequences from point clouds through recursive part decomposition, localized geometric contexts, and local CADQuery program synthesis.

Core training and layered recursive testing code for the CADRec grounding + contrastive learning experiment.

Main files

• cadgen_config.yaml: training configuration with contrastive learning and bbox grounding enabled.
• train.py, train.sh, traincadgen0311.py: training entrypoints.
• cadgen0311.py: model, collate, contrastive loss, and bbox grounding logic.
• cadgendataset0311.py: CAD dataset and sampling utilities.
• test_layer.py, recursive_bbox_infer0311.py, example1.py: layered recursive inference and testing utilities.
• yaml_config0311.py: YAML configuration loading and snapshot helpers.

Project website

The project page source lives in docs/ and uses plain HTML, CSS, and JavaScript, with no build step or third-party runtime dependencies. The live copy is published from shinodashx.github.io/dist/CADRec/ through that repository's existing GitHub Pages workflow. This keeps the URL at https://shinodashx.github.io/CADRec/ without requiring a separate Pages site for the code repository.

To preview locally, run python3 -m http.server 8080 --directory docs and open the local server in your browser. Edit docs/index.html for paper content, docs/style.css for styling, and docs/script.js for the accessible result tabs and citation-copy interaction. Video and optimized figures are in docs/assets/; the paper and full-resolution figures are in docs/files/.

To publish an update, copy the updated contents of docs/ into dist/CADRec/ in the shinodashx.github.io repository and push its main branch. Do not replace the personal homepage's other files.

The paper PDF is compiled from the provided camera-ready LaTeX sources using the current acmart class, with duplicated post-preamble publication declarations removed for compilation. Research content is unchanged. Figure previews come from the supplied PDFs; the pipeline uses the paper's final figure. The two supplied video files are identical, so the site includes one copy with English captions.

Citation

@article{song2026cadrec,
  title = {CADRec: Reconstructing a CAD Sequence Recursively with Localized Geometric Contexts},
  author = {Song, Haoxuan and Yang, Bingchen and Xiao, Jun and Jiang, Haiyong},
  journal = {ACM Transactions on Graphics},
  volume = {45},
  number = {6},
  articleno = {212},
  year = {2026},
  month = dec,
  doi = {10.1145/3842561}
}