SSP-GAN:Spectral Normalization Sphere-Guided 3D Point Cloud Generation

The proposed SSP-GAN is designed to tackle stability and convergence issues encountered during the training of three-dimensional point cloud generative adversarial networks (GANs). We introduce spectral normalization to the SP-GAN architecture, constraining the noise generation process of the generator to produce shapes that are more uniform and adhere closely to real distributions. Additionally, spectral normalization is applied to the discriminator to effectively limit its learning capacity, preventing overfitting to training data. Furthermore, we implemented a training strategy utilizing a two time-scale update rule, adjusting the update frequencies for both the generator and discriminator. The generator updates at a slower pace compared to the discriminator, facilitating convergence of the 3D point cloud GAN to a local Nash equilibrium. Experimental results demonstrate that our approach outperforms SP-GAN across various evaluation metrics, exhibiting improved convergence and significantly reduced training times. The generated 3D shapes are richer and more detailed, showcasing the efficacy of SSP-GAN.