Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DeepRAHT: Learning Predictive RAHT for Point Cloud Attribute Compression

Chunyang Fu, Tai Qin, Shiqi Wang, Zhu Li

TL;DR

DeepRAHT tackles lossy point cloud attribute compression by embedding a differentiable, end-to-end RAHT transform within a learning reconstruction, paired with a learnable predictive RAHT mechanism. It introduces a dyadic RAHT implemented via differentiable sparse convolutions, a learnable IDW-based prediction with grandfather-scale compensation, and a Laplace-based rate proxy for robust variable-rate coding through run-length coding. Empirical results show BD-BR gains over G-PCCv23 and superior performance to prior learning-based PCAC methods, while achieving fast encoding/decoding and strong robustness. The approach offers a practical, distortion-controllable framework for high-quality PCAC with potential extensions to LiDAR and dynamic point clouds.

Abstract

Regional Adaptive Hierarchical Transform (RAHT) is an effective point cloud attribute compression (PCAC) method. However, its application in deep learning lacks research. In this paper, we propose an end-to-end RAHT framework for lossy PCAC based on the sparse tensor, called DeepRAHT. The RAHT transform is performed within the learning reconstruction process, without requiring manual RAHT for preprocessing. We also introduce the predictive RAHT to reduce bitrates and design a learning-based prediction model to enhance performance. Moreover, we devise a bitrate proxy that applies run-length coding to entropy model, achieving seamless variable-rate coding and improving robustness. DeepRAHT is a reversible and distortion-controllable framework, ensuring its lower bound performance and offering significant application potential. The experiments demonstrate that DeepRAHT is a high-performance, faster, and more robust solution than the baseline methods. Project Page: https://github.com/zb12138/DeepRAHT.

DeepRAHT: Learning Predictive RAHT for Point Cloud Attribute Compression

TL;DR

DeepRAHT tackles lossy point cloud attribute compression by embedding a differentiable, end-to-end RAHT transform within a learning reconstruction, paired with a learnable predictive RAHT mechanism. It introduces a dyadic RAHT implemented via differentiable sparse convolutions, a learnable IDW-based prediction with grandfather-scale compensation, and a Laplace-based rate proxy for robust variable-rate coding through run-length coding. Empirical results show BD-BR gains over G-PCCv23 and superior performance to prior learning-based PCAC methods, while achieving fast encoding/decoding and strong robustness. The approach offers a practical, distortion-controllable framework for high-quality PCAC with potential extensions to LiDAR and dynamic point clouds.

Abstract

Regional Adaptive Hierarchical Transform (RAHT) is an effective point cloud attribute compression (PCAC) method. However, its application in deep learning lacks research. In this paper, we propose an end-to-end RAHT framework for lossy PCAC based on the sparse tensor, called DeepRAHT. The RAHT transform is performed within the learning reconstruction process, without requiring manual RAHT for preprocessing. We also introduce the predictive RAHT to reduce bitrates and design a learning-based prediction model to enhance performance. Moreover, we devise a bitrate proxy that applies run-length coding to entropy model, achieving seamless variable-rate coding and improving robustness. DeepRAHT is a reversible and distortion-controllable framework, ensuring its lower bound performance and offering significant application potential. The experiments demonstrate that DeepRAHT is a high-performance, faster, and more robust solution than the baseline methods. Project Page: https://github.com/zb12138/DeepRAHT.
Paper Structure (28 sections, 18 equations, 8 figures, 3 tables)

This paper contains 28 sections, 18 equations, 8 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional PCAC Methods
  4. Learning-based PCAC Methods
  5. Methodology
  6. Overview
  7. Transform Model
  8. Implementation
  9. Prediction Model
  10. IDW Prediction
  11. Prediction Compensation
  12. Entropy Coder
  13. Learning
  14. Experiments
  15. Experimental Setup
  16. ...and 13 more sections

Figures (8)

  • Figure 1: Overview of DeepRAHT. $A_0$ is the input attributes and $\hat{A}_0$ is the reconstructed attributes.
  • Figure 2: Dyadic RAHT decomposition.
  • Figure 3: Details of Haar module.
  • Figure 4: Prediction compensation module.
  • Figure 5: R-D curves averaged over the datasets.
  • ...and 3 more figures