MSG-Point-GAN: Multi-Scale Gradient Point GAN for Point Cloud Generation

The generative adversarial network (GAN) has recently emerged as a promising model. Its application in the image field been extensive, but there little research concerning point clouds.The combination of GAN and graph convolutional state-of-the-art method for generating clouds. However, is significant gap between generated cloud used training. In order to improve quality cloud, this study proposed multi-scale gradient (MSG-Point-GAN). training dynamic game process, we expected generation discrimination capabilities be symmetric, so that would more stable. Based on concept progressive growth, structure (MSG-GAN) stabilize process. discriminator part PointNet resolve problem disorder rotation cloud. could effectively determine authenticity This also analyzed optimization process objective function MSG-Point-GAN. experimental results showed MSG-Point-GAN was stable, superior other methods subjective vision. From perspective performance metrics, by real significantly smaller than methods. practical analysis, point-cloud classification improved about 0.2%, compared original network. provided stable framework generation. It can promote development point-cloud-generation technology.