Machine Perception Point Cloud Quality Assessment Via Vision Tasks

Abstract

LiDAR sensing is a newly developed 3D acquisition technology which is widely applied in auto-driving area. Different from the human perception point cloud, the generated 3D data is machine perception point clouds which are designed for specific vision tasks in realistic life, such as point cloud detection, segmentation and recognition. Therefore, instead of traditional subjective quality estimation, the quality assessment of machine perception point cloud is a new challenge. In this paper, we propose a machine perception point cloud quality assessment via various vision tasks, evaluating the point cloud quality based on the performance in vision tasks of different level of distorted point cloud. Firstly, we utilize the state-of-the-art point cloud compression algorithm to obtain the distorted point cloud. Then, we explore the potentials of distorted point clouds in detection and segmentation precision, comparing the results in different testing conditions. Finally, we propose the machine perception ROI based point cloud compression framework achieves notable performance on vision tasks result while do insignificant influence on PSNR.The experimental results illustrate the correspondence between point cloud quality and the performance in vision tasks, verifying the effectiveness of the proposed method.