SphereGAN: Sphere Generative Adversarial Network Based on Geometric Moment Matching and its Applications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) are powerful tools for data generation.
• Existing GANs face challenges in stable training and achieving high convergence rates.
• Measuring distances between probability distributions is crucial for GAN performance.

Purpose of the Study:

• To introduce SphereGAN, a novel GAN utilizing an integral probability metric on a hypersphere.
• To improve the stability and convergence rate of GAN training.
• To enhance the realism and accuracy of generated data, including images and 3D point clouds.

Main Methods:

• Developed SphereGAN, which measures probability distribution distances on a hypersphere.
• Employed a hypersphere-based objective function calculating distance as a half arc for stable training.
• Incorporated higher-order data information using multiple geometric moments for improved distance measurement accuracy.

Main Results:

• SphereGAN demonstrated stable training and a high convergence rate.
• The method achieved superior accuracy and convergence compared to state-of-the-art GANs.
• Quantitative and qualitative experiments on CIFAR-10, STL-10, LSUN bedroom, and ShapeNet datasets validated SphereGAN's effectiveness.

Conclusions:

• SphereGAN offers a robust and efficient approach to generative adversarial networks.
• The hypersphere-based metric and geometric moments significantly improve GAN performance.
• SphereGAN shows strong potential for unsupervised image and 3D point cloud generation tasks.