Large-scale point cloud contour extraction via 3D guided multi-conditional generative adversarial network

As one of the most important features for human perception, contours are widely used in many graphics and mapping applications. However, for large outdoor scale point clouds, contour extraction is considerably challenging due to the huge, unstructured and irregular point space, thus leading to massive failure for existing approaches. In this paper, to generate contours consistent with human perception for outdoor scenes, we propose, for the first time, 3D guided multi-conditional GAN (3D-GMcGAN), a deep neural network based contour extraction network for large scale point clouds. Specifically, two ideas are proposed to enable the network to learn the distributions of labeled samples. First, a parametric space based framework is proposed via a novel similarity measurement of two parametric models. Such a framework significantly compresses the huge point data space, thus making it much easier for the network to “remember” target distribution. Second, to prevent network loss in the huge solution space, a guided learning framework is designed to assist finding the target contour distribution via an initial guidance. To evaluate the effectiveness of the pro-posed network, we open-sourced the first, to our knowledge, dataset for large scale point cloud with contour annotation information. Experimental results demonstrate that 3D-GMcGAN efficiently generates contours for the data with more than ten million points (about several minutes), while avoiding ad hoc stages or parameters. Also, the proposed framework produces minimal outliers and pseudo-contours, as suggested by comparisons with the state-of-the-art approaches.