Rate-Reconfigurable Deep Point Cloud Compression With Perceptual Bit Allocation Optimization

Area of Science:

• Computer Vision
• Data Compression
• Machine Learning

Background:

• Conventional point cloud compression methods necessitate multiple models for varying bit rates.
• Existing approaches often fail to adequately address the differing rate requirements of geometry and attribute data.
• The need for efficient, adaptable, and perceptually optimized point cloud compression is critical.

Purpose of the Study:

• To propose an end-to-end Rate-Reconfigurable Deep Point Cloud Compression (RR-DPCC) framework.
• To enable arbitrary bit rate control using a single trained deep learning model.
• To enhance compression efficiency by jointly encoding geometry and attribute data with perceptual optimization.

Main Methods:

• Development of a Rate-Reconfigurable Deep Point Cloud Compression (RR-DPCC) framework incorporating on/off-line Perceptual Bit Allocation Optimization (PBAO-ON/OFF).
• Introduction of a one-stream network for joint geometry and attribute encoding.
• Implementation of a bitrate reconfigurable module and a rate allocation module for fine-grained control and optimized bit distribution.
• Derivation of rate-distortion models (R-α/β and D-α/β) for accurate rate control and bit allocation.

Main Results:

• The RR-DPCC achieves fine-grained bitrate control and allocation via a single trained model.
• Significant bit rate reductions were observed: -6.56% (PBAO-ON) and -4.90% (PBAO-OFF) compared to V-PCC.
• Further reductions of -18.68% (PBAO-ON) and -15.34% (PBAO-OFF) were achieved against Deep-JGAC.
• Substantial reductions in encoding/decoding time were reported: up to 98.38% (PBAO-ON) and 53.75% (PBAO-OFF) versus V-PCC.

Conclusions:

• The proposed RR-DPCC with PBAO-ON/OFF offers a unified solution for adaptable and efficient point cloud compression.
• The method achieves superior rate-distortion performance and significant computational savings.
• This approach advances the state-of-the-art in deep learning-based point cloud compression.