eduzhai > Applied Sciences > Engineering >

Rethinking Sampling in 3D Point Cloud Generative Adversarial Networks

  • Save

... pages left unread,continue reading

Document pages: 13 pages

Abstract: In this paper, we examine the long-neglected yet important effects of pointsampling patterns in point cloud GANs. Through extensive experiments, we showthat sampling-insensitive discriminators (e.g.PointNet-Max) produce shape pointclouds with point clustering artifacts while sampling-oversensitivediscriminators (e.g.PointNet++, DGCNN) fail to guide valid shape generation. Wepropose the concept of sampling spectrum to depict the different samplingsensitivities of discriminators. We further study how different evaluationmetrics weigh the sampling pattern against the geometry and propose severalperceptual metrics forming a sampling spectrum of metrics. Guided by theproposed sampling spectrum, we discover a middle-point sampling-aware baselinediscriminator, PointNet-Mix, which improves all existing point cloud generatorsby a large margin on sampling-related metrics. We point out that, though recentresearch has been focused on the generator design, the main bottleneck of pointcloud GAN actually lies in the discriminator design. Our work provides bothsuggestions and tools for building future discriminators. We will release thecode to facilitate future research.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×