2021 Wu DeepCAD Generative CAD Models

Deep generative models of 3D shapes have garnered significant attention, yet most generate discrete representations such as voxels and meshes. This study introduces the first 3D generative model that represents shapes as sequences of computer-aided design (CAD) operations, reflecting the user creation process crucial in various industrial applications. The unique sequential and irregular structures of CAD operations present challenges to existing models. By drawing parallels between CAD operations and natural language, we propose a CAD generative network leveraging the Transformer architecture. Our methodology includes the development of a novel CAD dataset containing 178,238 models and their accompanying CAD construction sequences, which has been made publicly available. We demonstrate the efficacy of our model for tasks such as shape autoencoding and random shape generation, showcasing its ability to produce diverse CAD designs. Various metrics, including coverage, minimum matching distance, and Jensen-Shannon divergence, are employed to evaluate the generated shapes against the reference dataset. Results indicate that our approach significantly advances the state of 3D shape generation by integrating the design process into generative modeling. Further research directions are anticipated based on this novel dataset and model framework.