Wasserstein distributional harvesting for highly dense 3D point clouds

Abstract

In this paper, we present a novel 3D point cloud harvesting method, which can harvest 3D points from an estimated surface distribution in an unsupervised manner (i.e., an input is a prior distribution). Our method outputs the surface distribution of a 3D object and samples 3D points from the distribution based on the proposed progressive random sampling strategy. The progressive sampling regards a prior distribution itself as a network input and uses a progressively increasing number of latent variables for training, which can diversify the coordinates of 3D points with fast convergence. Subsequently, our stochastic instance normalization transforms the implicit distribution into other distributions, which enables diverse shapes of 3D objects. Experimental results show that our method is competitive with other state-of-the-art methods. Our method can harvest an arbitrary number of 3D points, wherein the 3D object is represented in detail with highly dense 3D points or a part of it is described with partial sampling.