Weakly Supervised Point Cloud Upsampling via Optimal Transport

Existing learning-based methods usually train a point cloud upsampling model with synthesized, paired sparse-dense point clouds. However, the distribution gap between synthesized and real data limits the performance and generalization. To solve this problem, we innovatively regard the upsampling task as an optimal transport (OT) problem from sparse to dense point cloud. Further we propose PU-CycGAN, a cycle network that consists of a Densifier, Sparsifier and two discriminators. It can be directly trained for upsampling with unpaired real sparse point clouds, so that the distribution gap can be filled via the learning. Especially, quadratic Wasserstein distance is introduced for the stable training. Extensive experiments on both synthetic and real-scanned datasets validate the effectiveness and advantages in terms of distribution uniformity, underlying surface representation and applicability to real data.