SECAD-Net: Self-Supervised CAD Reconstruction by Learning Sketch-Extrude Operations

Reverse engineering CAD models from raw geometry is a classic but strenuousresearch problem. Previous learning-based methods rely heavily on labels due tothe supervised design patterns or reconstruct CAD shapes that are not easilyeditable. In this work, we introduce SECAD-Net, an end-to-end neural networkaimed at reconstructing compact and easy-to-edit CAD models in aself-supervised manner. Drawing inspiration from the modeling language that ismost commonly used in modern CAD software, we propose to learn 2D sketches and3D extrusion parameters from raw shapes, from which a set of extrusioncylinders can be generated by extruding each sketch from a 2D plane into a 3Dbody. By incorporating the Boolean operation (i.e., union), these cylinders canbe combined to closely approximate the target geometry. We advocate the use ofimplicit fields for sketch representation, which allows for creating CADvariations by interpolating latent codes in the sketch latent space. Extensiveexperiments on both ABC and Fusion 360 datasets demonstrate the effectivenessof our method, and show superiority over state-of-the-art alternativesincluding the closely related method for supervised CAD reconstruction. Wefurther apply our approach to CAD editing and single-view CAD reconstruction.The code is released at https://github.com/BunnySoCrazy/SECAD-Net.