CF-CAD: A Contrastive Fusion Network For 3D Computer-Aided Design Generative Modeling

Xueyang Li, Haotian Chen, Yunzhong Lou, Xiangdong Zhou

Abstract: Deep representation learning methods for Computer-Aided Design (CAD) generative modeling have received increasing attention recently. However, parametric CAD sequences as the key to construct 3D CAD models essentially contain the implicit geometry, which has not been concerned well. In this paper, we introduce Contrastive Fusion CAD (CF-CAD), a self-supervised multi-modal framework with a Transformer-based architecture. Our model enhances the correlation between the parametric CAD sequence and its geometry via additionally drawing 2D images into parametric CAD sequences with one shared codebook, which can optimize the latent representation of CAD models and empower it to achieve an interesting task of reversing images to parametric CAD sequences simultaneously. To further improve the alignment and uniformity of learned latent space, we also introduce a contrastive strategy to strike the well balance between parametric CAD sequences and 2D images. Extensive experiments on the commonly used benchmark datasets demonstrate the effectiveness of our CF-CAD for 3D CAD generative modeling.