Point cloud completion aims to infer missing regions of a point cloud, given an incomplete point cloud. Like image inpainting, in the 2D domain, point cloud completion offers a way to recreate an entire point cloud, given only a subset of the information. However, current applications study only synthetic datasets with artificial point removal, such as the Completion3D dataset. Although these datasets are valuable, they are an artificial problem set that we can not apply to real-world data. This paper draws a parallel between point cloud completion and occlusion reduction in aerial lidar scenes. We propose a crucial change in the hierarchical sampling using selforganizing maps to propose new points representing the scene in a reduced resolution. These new points are a weighted combination of the original set using spatial and feature information. A new set of proposed points is more powerful than simply sampling existing points. We demonstrate this sampling technique by replacing the farthest point sampling in the Skip-attention Network with Hierarchical Folding (SA-Net) and show a significant increase in the overall results using the Chamfers distance as our metric. We also show that we can use this sampling method in the context of any technique which uses farthest point sampling.
|