Deep G-PCC Geometry Preprocessing via Joint Optimization With a Differentiable Codec Surrogate for Enhanced...

Area of Science:

• Computer Vision
• Data Compression
• Geometric Modeling

Background:

• Geometry-based Point Cloud Compression (G-PCC) is an MPEG standard offering interoperability but lagging in compression efficiency compared to deep learning methods.
• Existing deep learning approaches often require significant modifications to established compression frameworks.

Purpose of the Study:

• To improve the rate-distortion performance of G-PCC without compromising its interoperability or computational flexibility.
• To develop a method that integrates deep learning enhancements with the existing G-PCC standard.

Main Methods:

• Proposed a compression-oriented point cloud voxelization network optimized with a differentiable G-PCC surrogate model.
• The surrogate model approximates the rate-distortion behavior of G-PCC, enabling end-to-end training.
• The voxelization network employs adaptive learning-based voxelization, global scaling, pruning, and point editing for optimization.

Main Results:

• Achieved an average 38.84% BD-rate reduction compared to standard G-PCC.
• The proposed method requires only the lightweight voxelization network to be added to the G-PCC encoder, with no decoder modifications.
• Inference introduces no computational overhead for end-users.

Conclusions:

• The integration of a learning-based voxelization network offers a practical enhancement for the G-PCC standard.
• This approach successfully bridges classical codecs and deep learning, improving compression efficiency while maintaining backward compatibility.
• The method is well-suited for real-world deployment scenarios requiring efficient point cloud compression.