Uy Mikaela Angelina ; Chang Yen-yu ; Sung Minhyuk ; Goel Purvi ; Lambourne Joseph ; Birdal Tolga ; Guibas Leonidas

Abstract

We propose Point2Cyl, a supervised network transforming a raw 3D point cloudto a set of extrusion cylinders. Reverse engineering from a raw geometry to aCAD model is an essential task to enable manipulation of the 3D data in shapeediting software and thus expand their usages in many downstream applications.Particularly, the form of CAD models having a sequence of extrusion cylinders-- a 2D sketch plus an extrusion axis and range -- and their booleancombinations is not only widely used in the CAD community/software but also hasgreat expressivity of shapes, compared to having limited types of primitives(e.g., planes, spheres, and cylinders). In this work, we introduce a neuralnetwork that solves the extrusion cylinder decomposition problem in ageometry-grounded way by first learning underlying geometric proxies.Precisely, our approach first predicts per-point segmentation, base/barrellabels and normals, then estimates for the underlying extrusion parameters indifferentiable and closed-form formulations. Our experiments show that ourapproach demonstrates the best performance on two recent CAD datasets, FusionGallery and DeepCAD, and we further showcase our approach on reverseengineering and editing.