Reproduce “CADFit: Precise Mesh-to-CAD Program Generation with Hybrid Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ghadi Nehme, Eamon Whalen, Faez Ahmed
CADFit reconstructs editable, executable parametric CAD programs from meshes by extracting sketch profiles, generating extrusion and revolution candidates, and selecting operations through CAD-kernel execution and geometric fit. On ABC Hard, a 300-shape subset defined by STEP face counts of 151 to 1,500, it reports IoU 0.662 versus Cadrille’s 0.438 and invalid ratio 0.000; this measures geometry, not recovered design intent.
Key insights
- Explicitly optimizing for geometric fidelity through an IoU-driven approach, validated by a CAD kernel, enables the precise recovery of complex parametric operations and guarantees program validity.
- A hybrid methodology, combining geometry-driven candidate generation with a learning-based sketch prior, effectively navigates the large, non-convex, and combinatorial search space of CAD reconstruction.
- Decoupling geometric perception from parametric reconstruction allows for robust operation across diverse input modalities, such as meshes and images, without requiring modality-specific retraining of the core reconstruction algorithm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.