Dupont Elona ; Cherenkova Kseniya ; Mallis Dimitrios ; Gusev Gleb ; Kacem Anis ; Aouada Djamila

Abstract

3D reverse engineering, in which a CAD model is inferred given a 3D scan of aphysical object, is a research direction that offers many promising practicalapplications. This paper proposes TransCAD, an end-to-end transformer-basedarchitecture that predicts the CAD sequence from a point cloud. TransCADleverages the structure of CAD sequences by using a hierarchical learningstrategy. A loop refiner is also introduced to regress sketch primitiveparameters. Rigorous experimentation on the DeepCAD and Fusion360 datasets showthat TransCAD achieves state-of-the-art results. The result analysis issupported with a proposed metric for CAD sequence, the mean Average Precisionof CAD Sequence, that addresses the limitations of existing metrics.