Self-Supervised Representation Learning for CAD Models

In product development, design decisions and processing steps are closely linked to the geometric shape of the product. Self-supervised representation learning methods make it possible to learn compact, versatile representations from a wide range of unlabeled data. These representations encode basic structures and features of the input data and can be used to solve application tasks in a data-efficient way. A key problem is the formulation of suitable self-supervised tasks that promote the learning of meaningful representations. While existing research in the field of computer vision primarily deals with pure point cloud data, the focus in this work is specifically on the use of CAD data.