Dynamic Point Cloud Compression with Shape and Density-based Volumetric Partitioning

A point cloud embodies objects or scenes by utilising unordered points representing geometric positions in 3D space along with attributes like colours and reflections, capturing the intricate details of natural forms. This unique capability has garnered significant attention across various applied fields, including virtual and augmented reality, while offering substantial advancements in training, safety, entertainment, and quality of life. Point clouds, especially dynamic point clouds, demand efficient compression due to their significant data volume. For example, consider a typical dynamic point cloud (a single person with low movement), which includes around 1 million points per frame and a frame rate of 30Hz; this would require more than 1.6 Gbps bandwidth without compression, which is more than 300 times a traditional full HD video.
The latest dynamic point cloud compression standard is video-based point cloud compression (V-PCC), which relies on projecting 3D input into 2D frames. The V-PCC divides the whole 3D point clouds into patches for the projection using computationally expensive and time-consuming processes. The patches are then projected onto layers of 2D planes to apply existing video coding techniques. Dividing an entire point cloud into patches all at once and in a single instance, followed by projection, can lead to the loss of critical data, such as some original points or their spatial relationships. This loss induces artifacts that adversely affect user perception. Moreover, packing irregular projected patches produces many unoccupied pixels in 2D frames, which increases the bitrate requirements of the 2D video coding and degrades overall image quality.
This thesis addresses the above mentioned problems of the V-PCC by partitioning point clouds into smaller parts to separate the regions based on the proximity of similar shapes and the density of the points. Therefore, this partitioning enhances our understanding of local information, allowing for more effective data management, capture, and retention. The proposed shape-based strategy can pack the points related to similar shapes into the same neighbouring area, thereby maintaining the proximity of the points and increasing the compression efficiency of 2D video coding. Density-based partitioning addresses original point loss because it can provide more patches where the density is high to reduce occluded points. This helps retain more original points during 2D projection and improves the image quality. This thesis also introduces a regular partitioning strategy to reduce the need to transmit overhead information about the positions of each partition. Additionally, this regular partitioning avoids costly and time-consuming processes during patch generation, thereby reducing time complexity. While reducing time complexity and data loss, the regular partitioning method decreases the size of the 2D frames since the projected patches are mostly regular and can be packed more closely, thereby decreasing unoccupied pixels. Finally, this thesis has proposed a hierarchical partitioning technique that can adapt to varying point cloud densities and non-planar surfaces for further improvement of the rate-distortion performance. Partitioning from different positions in the second layer is also proposed to improve patch boundary continuity, surpassing V-PCC's approach and effectively increasing the image quality. The compact 2D maps generated by the proposed methods allow two layers to be combined in a single map for further improving compression performance.
The proposed methods outperform current techniques, as proved by quantitative and qualitative analyses through extensive experiments on benchmark datasets. This thesis addresses critical challenges in dynamic point cloud compression. It presents innovative solutions to enhance compression efficiency, reduce data loss, and improve dynamic point cloud compression. Significantly advancing the field, this work is a comprehensive framework with valuable insights for real-time and broadcasting applications.