PCCRender: Joint Learning of Point Cloud Compression and Gaussian Splatting Rendering

Area of Science:

• Computer Vision
• Computer Graphics
• Data Compression

Background:

• Point clouds are vital 3D data for VR/AR and digital museums.
• Existing compression methods often compromise 2D or 3D visual quality.
• Demand for high-quality rendering and compression with lower bandwidth is increasing.

Purpose of the Study:

• Introduce PCCRender, an end-to-end framework for high-quality point cloud attribute preservation and rendering.
• Address challenges of bandwidth limitations and variable data density in point cloud processing.
• Achieve superior 2D and 3D visual quality alongside efficient compression.

Main Methods:

• Developed Geometry-Invariant Rate Adjustment (GIRA) for efficient variable rate compression.
• Introduced Uneven Four-Group Context Model (UFCM) to accelerate decoding speed.
• Designed Voxel to Gaussian Primitive Converter (V2G-Converter) for differentiable rendering.

Main Results:

• PCCRender achieves state-of-the-art performance in point cloud compression and rendering.
• Demonstrated 10.58% BD-Rate gain in attribute compression compared to GPCC v23 (GS Render).
• Achieved 1.5 dB BD-PSNR increase in 2D rendered visual quality.

Conclusions:

• PCCRender offers a practical solution for next-generation point cloud applications.
• The framework supports variable rate and free viewpoint rendering.
• Achieves a balance between compression efficiency, rendering quality, and attribute fidelity.