Deep Learning Based Point Cloud Processing and Compression
Metadata[+] Show full item record
A point cloud is a 3D data representation that is becoming increasingly popular. Recent significant advances in 3D sensors and capturing techniques have led to a surge in the usage of 3D point clouds in virtual reality/augmented reality (VR/AR) content creation, as well as 3D sensing for robotics, smart cities, telepresence, and automated driving applications. With an increase in point cloud applications and improved capturing technologies, we now have high-resolution point clouds with millions of points per frame. However, due to the large size of a point cloud, efficient techniques for the transmission, compression, and processing of point cloud content are still widely sought. This thesis addresses multiple issues in the transmission, compression, and processing pipeline for point cloud data. We employ a deep learning solution to process 3D dense as well as sparse point cloud data for both static as well as dynamic contents. Employing deep learning on point cloud data which is inherently sparse is a challenging task. We propose multiple deep learning-based frameworks that address each of the following problems: Point Cloud Compression Artifact Removal. V-PCC is the current state-of-the-art for dynamic point cloud compression. However, at lower bitrates, there are unpleasant artifacts introduced by V-PCC. We propose a deep learning solution for V-PCC artifact removal by leveraging the direction of projection property in V-PCC to remove quantization noise. Point Cloud Geometry Prediction. The current point cloud lossy compression and processing techniques suffer from quantization loss which results in a coarser sub-sampled representation of the point cloud. We solve the problem of points lost during voxelization by performing geometry prediction across spatial scales using deep learning architecture. Point Cloud Geometry Upsampling. Loss of details and irregularities in point cloud geometry can occur during the capturing, processing, and compression pipeline. We present a novel geometry upsampling technique, PU-Dense, which can process a diverse set of point clouds including synthetic mesh-based point clouds, real-world high-resolution point clouds, real-world indoor LiDAR scanned objects, as well as outdoor dynamically acquired LiDAR-based point clouds. Dynamic Point Cloud Interpolation. Dense photorealistic point clouds can depict real-world dynamic objects in high resolution and with a high frame rate. Frame interpolation of such dynamic point clouds would enable the distribution, processing, and compression of such content. We also propose the first point cloud interpolation framework for photorealistic dynamic point clouds. Inter-frame Compression for Dynamic Point Clouds. Efficient point cloud compression is essential for applications like virtual and mixed reality, autonomous driving, and cultural heritage. We propose a deep learning-based inter-frame encoding scheme for dynamic point cloud geometry compression. In each case, our method achieves state-of-the-art results with significant improvement to the current technologies.
Table of Contents
Introduction -- Point cloud compression artifact removal -- Point cloud geometry prediction -- PU-Dense: sparse tensor-based point cloud geometry upsampling -- Dynamic point cloud interpolation -- Inter-frame compression for dynamic point cloud geometry coding
Ph.D. (Doctor of Philosophy)