3D Shape Generation via Variational Autoencoder with Signed Distance Function Relativistic Average Generative Adversarial Network

3D shape generation is widely applied in various industries to create, visualize, and analyse complex data, designs, simulations. Typically, uses a large dataset of shapes as the input. This paper proposes variational autoencoder with signed distance function relativistic average generative adversarial network, referred 3D-VAE-SDFRaGAN, for from 2D input images. Both network (GAN) (VAE) algorithms are typical used generate realistic shapes. However, it very challenging train stable model using VAE-GAN. an efficient approach stabilize training process VAE-GAN high-quality A mesh-based first generated representation by feeding single image into 3D-VAE-SDFRaGAN network. The maintain inside–outside information implicit surface representation. In addition, discriminator loss employed function. polygon mesh surfaces then produced via marching cubes algorithm. proposed evaluated ShapeNet dataset. experimental results indicate notable enhancement qualitative performance, evidenced visual comparison samples, well quantitative performance evaluation chamfer metric. achieves score 0.578, demonstrating superior compared existing state-of-the-art models.