Generative Adversarial Network with U-Net Type Generators for Complementing 3D Point Cloud Data of the Surrounding Environment and New Metric for Evaluation of Completion Results

Abstract

This paper proposes an generative adversarial network with U-Net type generators to complement large-scale 3D point cloud data obtained by measuring the surrounding environment. Conventional 3D point cloud completion methods based on deep learning have a problem that the spatial computational complexity becomes enormous when processing large 3D point cloud data, making them difficult to implement in a general-purpose computer environment. The proposed network has been devised to solve this problem. Several distance measures have been used in the past to evaluate the performance of 3D point cloud data completion, but they simply calculate the distance between two point clouds without considering the correspondence between the two points. Therefore, it is not possible to calculate a similarity that reflects whether a point with respect to an input point is generated in the neighborhood of the input point or not. In this paper, we also propose a new criterion to evaluate the performance of complementing 3D point clouds obtained from mechanical scanning type LIDAR, and use this criterion as a loss function when training the proposed network.