Attribute Artifacts Removal for Geometry-Based Point Cloud Compression.

Area of Science:

• Computer Vision
• Data Compression
• Geometric Processing

Background:

• Geometry-based Point Cloud Compression (G-PCC) offers high efficiency but suffers from attribute compression artifacts, particularly at low bitrates.
• Existing methods struggle to effectively mitigate these artifacts, impacting the visual fidelity of point clouds.

Purpose of the Study:

• To develop an effective method for removing attribute compression artifacts in G-PCC.
• To enhance the visual quality of point clouds compressed using G-PCC, especially under challenging low-bitrate conditions.

Main Methods:

• A Multi-Scale Graph Attention Network (MS-GAT) was proposed, utilizing Chebyshev graph convolutions for attribute feature extraction.
• A multi-scale scheme was employed to capture both short- and long-range correlations between points.
• Quantization step per point and a weighted graph attentional layer were incorporated to address adaptive quantization and artifact severity.

Main Results:

• The MS-GAT method achieved an average BD-rate reduction of 9.74% compared to Predlift and 10.13% compared to RAHT.
• Subjective evaluations confirmed a significant reduction in visual artifacts, including color shifting, blurring, and quantization noise.
• This represents the first attribute artifact removal method specifically designed for G-PCC.

Conclusions:

• The proposed MS-GAT effectively removes attribute compression artifacts in G-PCC.
• The method offers substantial objective (BD-rate reduction) and subjective (visual quality) improvements.
• This work advances the field of point cloud compression by addressing a critical limitation of G-PCC.