Point Cloud Compression with Implicit Neural Representations: A Unified Framework

TL;DR

The paper addresses the challenge of efficiently compressing unstructured point clouds by introducing a unified framework that jointly encodes geometry and attributes using two neural implicit representations. Geometry is captured by an occupancy network $f_{oldsymbol{ heta}}$ and attributes by a color network $g_{oldsymbol{ ho}}$, operating on a voxelized space with non-empty-cube pruning and per-cloud training. After learning, network parameters are quantized and entropy-coded with DeepCABAC, and decoding reconstructs the point cloud from the quantized parameters and auxiliary data. Empirically, the approach delivers state-of-the-art rate-distortion performance compared to the latest G-PCC octree baseline for geometry and joint geometry-attribute coding, while exhibiting strong universality across unseen clouds. This framework opens avenues for robust PCC, with potential extensions to dynamic scenes through shared neural representations across frames.

Abstract

Point clouds have become increasingly vital across various applications thanks to their ability to realistically depict 3D objects and scenes. Nevertheless, effectively compressing unstructured, high-precision point cloud data remains a significant challenge. In this paper, we present a pioneering point cloud compression framework capable of handling both geometry and attribute components. Unlike traditional approaches and existing learning-based methods, our framework utilizes two coordinate-based neural networks to implicitly represent a voxelized point cloud. The first network generates the occupancy status of a voxel, while the second network determines the attributes of an occupied voxel. To tackle an immense number of voxels within the volumetric space, we partition the space into smaller cubes and focus solely on voxels within non-empty cubes. By feeding the coordinates of these voxels into the respective networks, we reconstruct the geometry and attribute components of the original point cloud. The neural network parameters are further quantized and compressed. Experimental results underscore the superior performance of our proposed method compared to the octree-based approach employed in the latest G-PCC standards. Moreover, our method exhibits high universality when contrasted with existing learning-based techniques.

Figures (6)

• Figure 1: Overview of the proposed method. (a) The training and inference procedure of the networks, indicated by dashed lines and solid lines respectively. (b) The coding procedure.
• Figure 2: The detailed structure for both networks.
• Figure 3: Rate-distortion curves of different geometry compression methods, measured by D1 PSNR.
• Figure 4: Rate-distortion curves of the proposed method and G-PCC for joint geometry and attribute compression, measured by D1 PSNR and Y PSNR.
• Figure 5: Ablation studies on longdress. (a) Performance of attribute compression with different $L$. (b) Performance of geometry compression when $\beta= 0.1, 0.5, \zeta$. (c) Reconstructed geometry distortion (Left) and scaling ratio (Right) with regard to threshold $\tau$ at different bit rates. The optimal ratios are indicated by dashed lines.