Blind Quality Enhancement for G-PCC Compressed Dynamic Point Clouds
Researchers' blind enhancement model improves compressed 3D video quality by 0.5 dB PSNR, cutting bitrate needs by 20%.
A research team from City University of Hong Kong and other institutions has developed the first AI model capable of enhancing the quality of compressed 3D point cloud videos without needing to know how heavily they were compressed. Traditional methods require training separate models for each compression level, making them impractical for real-world use where compression settings are unknown. The new Blind Quality Enhancement (BQE) model solves this by using a two-branch architecture that extracts features across multiple distortion levels and adaptively fuses them based on estimated quality.
The BQE model specifically targets Geometry-based Point Cloud Compression (G-PCC), the MPEG standard for 3D data. It works by analyzing consecutive frames, using a motion compensation module to align them temporally, and then applying a cross-attention mechanism to understand temporal correlations. A quality estimation module predicts weighting distributions to guide the fusion of hierarchical features extracted at different compression levels. When tested on the G-PCC reference software (TMC13v28), BQE achieved significant improvements: average PSNR gains of 0.535 dB for luminance and around 0.4 dB for chrominance components, while reducing bitrate requirements (BD-rate) by approximately 20% across all components.
This breakthrough addresses a critical limitation in 3D content distribution. The model's ability to handle unknown compression levels with a single architecture makes it deployable in practical scenarios like streaming volumetric video for VR/AR, autonomous vehicle data transmission, and cultural heritage preservation. By improving visual quality while reducing bandwidth needs, BQE could accelerate adoption of high-quality 3D experiences across consumer and industrial applications.
- First blind quality enhancement model for G-PCC compressed point clouds that doesn't require prior knowledge of distortion level
- Achieves 0.535 dB PSNR improvement and 20.5% BD-rate reduction (bitrate savings) for compressed 3D video
- Uses adaptive feature fusion architecture with temporal correlation analysis across consecutive frames
Why It Matters
Enables higher quality 3D streaming for VR/AR and autonomous systems while using less bandwidth, solving a key deployment bottleneck.