Twins-Mix: Self Mixing in Latent Space for Reasonable Data Augmentation of 3D Computer-Aided Design Generative Modeling

Deep generative modeling on parametric Computer-Aided Design (CAD) data has great potentials in various industrial scenarios. However, the reasonable data augmentation for parametric CAD has not been solved yet, which limits the performance of the deep learning on parametric CAD data. Unlike images or language, parametric CAD data involves dual-aspects: command sequence and geometric shape. Hence, most previous data augmentation methods are unsuitable for this problem. To address this issue, we propose a novel mix-based augmentation method, namely Twins-Mix, keeping a well balance between the diversity and the validity of parametric CAD sequences, which significantly boost the performance of CAD generative modeling. Comprehensive experiments are conducted on the commonly used benchmark datasets, i.e., Fusion 360 and DeepCAD. The experimental results demonstrate that our model exceeds other comparable augmentations on CAD generative modeling significantly, especially in increasing the valid 3D shape construction ratio by at least more than 4%.