Com-PCQA: No-Reference Point Cloud Quality Assessment via Complex-Valued Feature Learning

Area of Science:

• Computer Vision
• Immersive Media Technologies
• Signal Processing

Background:

• Point Cloud Quality Assessment (PCQA) is vital for immersive media, but current methods struggle with complex visual cues.
• Existing real-valued PCQA techniques fail to capture coupled geometric and perceptual information effectively.
• Human evaluation is costly, and optimizing compression pipelines requires accurate quality metrics.

Purpose of the Study:

• To propose Com-PCQA, a novel no-reference PCQA framework utilizing complex-valued feature learning.
• To enhance the accuracy and robustness of point cloud quality assessment.
• To provide a framework that effectively models both geometric and perceptual aspects of visual quality.

Main Methods:

• A Hilbert dual-stream module converts point clouds and images into complex analytic signals for joint modeling.
• A complex amplitude-phase attention (CAPA) module separates and integrates geometric (amplitude) and detail (phase) features.
• An adversarial joint scoring module employs adversarial and collaborative learning for multi-modal, multi-scale representation calibration.

Main Results:

• Com-PCQA demonstrates state-of-the-art correlations with subjective quality scores.
• The proposed method consistently outperforms existing PCQA techniques across multiple datasets.
• Experiments validate the effectiveness and robustness of the complex-valued approach.

Conclusions:

• Com-PCQA offers a significant advancement in no-reference point cloud quality assessment.
• The complex-valued feature learning approach effectively captures essential visual quality cues.
• The framework shows promise for optimizing immersive media pipelines and enhancing user perception.