CADFit: Precise Mesh-to-CAD Program Generation with Hybrid Optimization

Despite advances in CAD technology, recovering parametric CAD construction sequences from geometric inputs like meshes remains a challenge in design and manufacturing. This paper introduces CADFit, a hybrid optimization-based framework that reconstructs complex, editable CAD sequences from meshes by incrementally fitting parametric operations and validating with geometric feedback. CADFit differentiates itself by performing an Intersection-over-Union (IoU)-driven optimization over structured CAD programs and incorporates various operations such as extrusions and fillets. Through extensive experiments with benchmark datasets, CADFit is shown to outperform existing mesh-to-CAD methods in both volumetric Intersection-over-Union and Chamfer Distance metrics, while significantly lowering the Invalid Ratio for reconstructed programs, especially with complex designs. Moreover, a multimodal pipeline that allows complete reconstruction from images is proposed, merging image-based geometry reconstruction with CADFit's capabilities. By facilitating the accurate reconstruction of higher-complexity CAD models, this work underlines a significant step towards generating richer datasets and enhances future learning-based approaches in CAD reverse engineering.