Voxel-based Point Cloud Geometry Compression with Space-to-Channel Context.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Data Compression

Background:

• Voxel-based methods are efficient for point cloud geometry compression but have limitations with receptive fields, especially for high-bit-depth data.
• Upsampling operations in current methods restrict their ability to capture long-range dependencies in point cloud data.

Purpose of the Study:

• To develop an advanced context model for enhanced point cloud geometry compression.
• To overcome the receptive field limitations of existing voxel-based compression techniques.
• To improve compression efficiency and reduce computational complexity for diverse point cloud types.

Main Methods:

• Introduced a stage-wise Space-to-Channel (S2C) context model for dense and low-level sparse point clouds, using channel-wise autoregression.
• Proposed a level-wise S2C context model for high-level sparse point clouds, integrating Geometry Residual Coding (GRC) for cross-level prediction.
• Utilized the spherical coordinate system and enhanced GRC with a Residual Probability Approximation (RPA) module.

Main Results:

• The S2C context model achieved significant bit savings compared to state-of-the-art methods.
• Reconstruction quality was maintained or improved with the proposed compression techniques.
• Computational complexity was reduced, demonstrating practical efficiency gains.

Conclusions:

• The novel S2C context model effectively addresses limitations in point cloud geometry compression.
• The proposed methods offer a superior balance of compression efficiency, reconstruction quality, and computational cost.
• This work advances the field of point cloud compression, particularly for dense and sparse data representations.