In this study the effectiveness of generative models like Variational Auto Encoders (VAE) for generation of con- strained compound shapes in the form of point cloud data was explored. Majority of generative methods in the literature have tackled generative approaches for a shape as a whole. But most of the engineering applications require shapes with multiple constituent part geometries/designs e.g. flow valves, vehicle bod- ies, tire assemblies. SDM-NET has tackled this problem using VAE models. But we observed that for our engineering design generative model use case, it has limitations in tackling intricate inter-part relations.
Our objective is to train an end-to-end generative point cloud model to explore latent space by respecting part connection constraints as well as maintaining variety in producing newer shapes. We have developed PointNET based permutation invariant generative model for unordered point cloud data with multiple improvements. We present results with VAE as generative model, but many of our proposed suggestions are applicable to other generative models like GAN and diffusion models as well. We have integrated self-attention mechanism, structural embeddings to capture inter-part relations, balancing of reconstruction and generative capabilities which are explained in detail. Due to the confidentiality of actual design shapes, we have presented our research work with synthetic data generated with basic geometric shapes.
Access The Research Paper:
Lattice | Vol 5 Issue 1₹1,657.00