Attention Focused Generative Network for Reducing Self-Occlusions in Aerial LiDAR
Nina M. Singer, Vijayan K. Asari, Theus H. Aspiras, Jonathan Schierl, Andrew Stokes, Brett L. Keaffaber, Andre J. Van Rynbach, Kevin Decker, David J. Rabb · 2021
This work explores the problem of re-ducing self-occlusion in aerial lidar. We introduce an entirely new dataset, DALES Viewpoints which uses a combination of synthetic and real-world data to represent aerial lidar scenes with different levels of occlusions. Our overall goal is to transform these occluded point clouds into a more visually complete representation of the same scene. We also propose a method called Channel Learned Downsampling (CLD). This downsampling method can be used as a drop-in replacement for any sampling method and uses a channel attention mechanism to select points based on their features instead of relying entirely on their spatial information. We show that this sampling method outperforms other sampling methods when used with a state-of-the-art point cloud completion network.