Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diff-PCC: Diffusion-based Neural Compression for 3D Point Clouds

Kai Liu, Kang You, Pan Gao

TL;DR

This work introduces Diff-PCC, a diffusion-based lossy point cloud compression framework that employs a dual-space latent encoding to extract complementary low- and high-frequency shape information and a diffusion-based generator that denoises noisy point clouds under the guidance of these latents. By integrating a hyperprior-driven rate model and a rate-distortion objective, Diff-PCC achieves superior compression performance compared to G-PCC and recent learning-based methods, with substantial BD-PSNR gains at ultra-low bitrates and improved perceptual quality. The method advances neural point cloud compression by addressing the limitations of Gaussian priors in VAEs and leveraging diffusion models for high-fidelity reconstruction, while also introducing architecture elements like AdaLN-based conditioning and cross-frequency feature fusion. Although effective, the approach incurs higher coding complexity and currently targets smaller-scale point clouds, pointing to future work in acceleration and scalability for broader 3D workloads.

Abstract

Stable diffusion networks have emerged as a groundbreaking development for their ability to produce realistic and detailed visual content. This characteristic renders them ideal decoders, capable of producing high-quality and aesthetically pleasing reconstructions. In this paper, we introduce the first diffusion-based point cloud compression method, dubbed Diff-PCC, to leverage the expressive power of the diffusion model for generative and aesthetically superior decoding. Different from the conventional autoencoder fashion, a dual-space latent representation is devised in this paper, in which a compressor composed of two independent encoding backbones is considered to extract expressive shape latents from distinct latent spaces. At the decoding side, a diffusion-based generator is devised to produce high-quality reconstructions by considering the shape latents as guidance to stochastically denoise the noisy point clouds. Experiments demonstrate that the proposed Diff-PCC achieves state-of-the-art compression performance (e.g., 7.711 dB BD-PSNR gains against the latest G-PCC standard at ultra-low bitrate) while attaining superior subjective quality. Source code will be made publicly available.

Diff-PCC: Diffusion-based Neural Compression for 3D Point Clouds

TL;DR

This work introduces Diff-PCC, a diffusion-based lossy point cloud compression framework that employs a dual-space latent encoding to extract complementary low- and high-frequency shape information and a diffusion-based generator that denoises noisy point clouds under the guidance of these latents. By integrating a hyperprior-driven rate model and a rate-distortion objective, Diff-PCC achieves superior compression performance compared to G-PCC and recent learning-based methods, with substantial BD-PSNR gains at ultra-low bitrates and improved perceptual quality. The method advances neural point cloud compression by addressing the limitations of Gaussian priors in VAEs and leveraging diffusion models for high-fidelity reconstruction, while also introducing architecture elements like AdaLN-based conditioning and cross-frequency feature fusion. Although effective, the approach incurs higher coding complexity and currently targets smaller-scale point clouds, pointing to future work in acceleration and scalability for broader 3D workloads.

Abstract

Stable diffusion networks have emerged as a groundbreaking development for their ability to produce realistic and detailed visual content. This characteristic renders them ideal decoders, capable of producing high-quality and aesthetically pleasing reconstructions. In this paper, we introduce the first diffusion-based point cloud compression method, dubbed Diff-PCC, to leverage the expressive power of the diffusion model for generative and aesthetically superior decoding. Different from the conventional autoencoder fashion, a dual-space latent representation is devised in this paper, in which a compressor composed of two independent encoding backbones is considered to extract expressive shape latents from distinct latent spaces. At the decoding side, a diffusion-based generator is devised to produce high-quality reconstructions by considering the shape latents as guidance to stochastically denoise the noisy point clouds. Experiments demonstrate that the proposed Diff-PCC achieves state-of-the-art compression performance (e.g., 7.711 dB BD-PSNR gains against the latest G-PCC standard at ultra-low bitrate) while attaining superior subjective quality. Source code will be made publicly available.
Paper Structure (20 sections, 15 equations, 4 figures, 2 tables)

This paper contains 20 sections, 15 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Our Approach
  4. Contribution
  5. Related Work
  6. Point Cloud Compression
  7. Diffusion Models for Point Cloud
  8. Method
  9. Preliminaries
  10. DIFF-PCC
  11. Overview
  12. Dual-Space Latent Encoding
  13. Diffusion-based Generator
  14. Training Objective
  15. Experiments
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Diff-PCC pipeline. $X_{t}$ and $\bar{X}_{t}$ represents the $t$th original point cloud and noisy point cloud, respectively; $p$ refers to the forward process and $q$ refers to the reverse process; $N(0,\boldsymbol{I})$ means the pure noise. Entropy model and arithmetic coding is omitted for a concise explanation.
  • Figure 2: Detailed Structure of the Utilized Compressor and Generator. $y_l$ and $y_h$ refer to the low-frequency shape latent and high-frequency detail latent, respectively; $z$ means hyperprior latent; $Q$ refers to the quantization; AE and AD represents the arithmetic encoding and decoding.
  • Figure 3: Rate-distortion curves for performance comparison. From left to right: ShapeNet, ModelNet10, and ModelNet40 dataset.
  • Figure 4: Subjective quality comparison. Example point clouds are selected from the ShapeNet dataset, each with 2k points.