Comparison on Upscaling and Upsampling Methods of Underwater Sonar Point Clouds

Abstract

We are developing a remotely controlled unmanned system for harbor infrastructure construction. In order to construct underwater sonar point cloud dataset for automatic tetrapod detection, we measured point clouds of 30 and 120 kg tetrapods in this study. First, the measured sonar point clouds of 30 kg tetrapod were upscaled by 1.58 times after subtracting them from the centroid of the tetrapod point clouds. And then we applied two upsampling methods, adversarial residual graph convolution network (AR-GCN) and point cloud library (PCL) to increase the density of the upscaled sonar data of 30 kg tetrapod. The performance of the upscaled and upsampled results were evaluated by Chamfer distance and earth mover’s distance. The new ARGCN trained with TTP dataset showed better qualitative and quantitative performances than PCL and AR-GCN trained with conventional dataset.