GATF-PCQA: A Graph Attention Transformer Fusion Network for Point Cloud Quality Assessment

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Point cloud quality assessment is challenging due to high dimensionality and irregular 3D data structures.
• Aligning objective quality predictions with human perception is crucial but difficult.

Purpose of the Study:

• To develop a novel graph-based learning architecture for accurate point cloud quality assessment.
• To integrate perceptual features with advanced graph neural networks to mimic human judgment.

Main Methods:

• Extracted key perceptual features (curvature, saliency, color) to capture geometric and visual distortions.
• Constructed a graph representation with perceptual clusters as nodes and feature similarities as edges.
• Employed a Graph Attention Network Transformer Fusion (GATF) module for feature refinement and a Graph Convolutional Network (GCN) for quality score regression.

Main Results:

• The proposed method demonstrated high correlation with human subjective quality scores.
• Achieved superior performance compared to existing state-of-the-art metrics on benchmark datasets (ICIP2020, WPC, SJTU-PCQA).

Conclusions:

• The novel graph-based learning architecture effectively models perceptual mechanisms for quality judgment.
• The method offers a robust solution for objective point cloud quality assessment aligned with human perception.