PCE-GAN: A Generative Adversarial Network for Point Cloud Attribute Quality Enhancement Based on Optimal Transport

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Data Compression

Background:

• Point cloud compression reduces data size but often degrades reconstruction quality.
• Existing methods prioritize data fidelity over perceptual quality, which is crucial for human visual interpretation.

Purpose of the Study:

• To develop an advanced quality enhancement technique for compressed point clouds.
• To simultaneously optimize both data fidelity and perceptual quality using a novel generative adversarial network.

Main Methods:

• Proposed a generative adversarial network for point cloud quality enhancement (PCE-GAN) based on optimal transport theory.
• Generator includes Local Feature Extraction (LFE) using dynamic graphs and attention, and Global Spatial Correlation (GSC) using transformers.
• Discriminator enforces distribution matching between enhanced and original point clouds.

Main Results:

• PCE-GAN achieved state-of-the-art performance in point cloud quality enhancement.
• Demonstrated significant BD-rate improvements (e.g., -19.2% vs. PredLift) when applied to geometry-based point cloud compression (G-PCC).
• Subjective evaluations showed enhanced texture clarity, smoother color transitions, and finer detail preservation.

Conclusions:

• PCE-GAN effectively enhances point cloud quality by balancing data fidelity and perceptual metrics.
• The proposed method offers a significant improvement over existing compression and enhancement techniques.
• PCE-GAN shows promise for applications requiring high-quality 3D data reconstruction.