Synthesizing Dense and Colored 3D Point Clouds for Training Deep Neural Networks

3D point clouds are a compact homogeneous representation that have the ability to cap- ture intricate details of the environment. They are useful for a wide variety of applications. For example, point clouds can be sampled from the mesh of manually designed objects to use as synthetic data for training deep learning networks. However, the geometry and tex- ture of these point clouds is bounded by the resolution of the modeled objects. To facilitate learning with synthetic 3D point clouds, we present a novel conditional generative adver- sarial network that creates dense point clouds, with color, in an unsupervised manner. The difficulty of capturing intricate details at high resolutions is handled by a point transformer that progressively grows the network through the use of graph convolutions. Every training iteration evolves a point vector into a point cloud. Experimental results show that our net- work is capable of learning a 3D data distribution and produces colored point clouds with fine details at multiple resolutions.