A Synthetic CAD Models Dataset for Deep Learning

3D reconstruction is a significant research topic in the field of Computer-Aided Design (CAD), which is used to recover editable CAD models from original shapes, including point clouds, voxels, meshes, or boundary representations (B-rep). In recent years, deep learning methods (DL) have exhibited significant potential in the field of CAD, but there is currently a lack of large-scale CAD model datasets that support deep learning for parametrized feature-based modeling. Therefore, we synthesized a large-scale dataset containing one million CAD designs to provide labeled CAD model data for supervised training in relevant deep learning tasks.We employed a random synthesis algorithm to generate CAD models and documented the corresponding feature-based modeling processes, including principal primitives (cuboids, prisms, cylinders, cones, and spheres) and detail features (slots, semi-circular slots, through-holes, steps, fillets, chamfers, etc.). For each CAD model in the dataset, we provided four types of data files: a STEP file recording the 3D shape of the CAD model (compliant with ISO 10303-21 and GB/T 16656 standards), a JSON file recording the parametric feature modeling process (compliant with ECMA-404 and ISO/IEC 21778:2017 standards), a B-Rep graph representation file for CAD models (with a .bin suffix, in the format of the open-source graph neural network framework Deep Graph Library), and three-dimensional isometric side views of the CAD model.