Point cloud geometry compression using neural implicit representations

In recent years, the increasing prominence of 3D point clouds in various applications has led to an escalating need for efficient storage and transmission methods. The sheer size of these point cloud datasets presents challenges in rendering, transmission, and general usability. This thesis introduces a novel approach to point cloud geometry compression leveraging neural implicit representations, specifically through the use of a DiGS network model. By training this model on a single point cloud, we achieve a compact neural representation of its geometry. Notably, this representation allows for the reconstruction of the point cloud with an arbitrary resolution. After training a reconstructing network, dynamic quantization is applied on the trained weights, significantly reducing its overall bitrate without strongly compromising the quality of the reconstructed point cloud. A dequantization is then used to rebuild a high-fidelity representation of the original point cloud. Our experimental results demonstrate the efficacy of this approach in terms of compression ratios and reconstruction quality, assessed using PSNR relative to the bitrate. This research provides a promising direction for efficient point cloud geometry storage and transmission, addressing some of the growing demands of the 3D data era.