Inter-Frame Compression for Dynamic Point Cloud Geometry Coding.

Area of Science:

• Computer Vision
• Data Compression
• Deep Learning

Background:

• Dynamic point cloud geometry compression is crucial for emerging technologies like virtual and augmented reality, autonomous driving, and digital preservation.
• Existing compression methods, including Geometry-based Point Cloud Compression (G-PCC) and Video-based Point Cloud Compression (V-PCC), face challenges in efficiently handling the temporal redundancy of dynamic point clouds.

Purpose of the Study:

• To develop an efficient deep learning-based inter-frame encoding scheme for dynamic point cloud geometry compression.
• To improve compression performance by leveraging temporal correlations between consecutive point cloud frames.

Main Methods:

• A novel feature space inter-prediction network using sparse convolutions and hierarchical multiscale 3D feature learning is proposed.
• Motion compensation in the feature domain is employed to predict the current frame's latent representation from the previous frame.
• A learned probabilistic factorized entropy model is used to compress the residual features.

Main Results:

• The proposed method achieved significant rate reductions compared to state-of-the-art MPEG standards: over 88% BD-Rate reduction against G-PCCv20 Octree, over 56% against G-PCCv20 Trisoup, over 62% against V-PCC intra-frame, and over 52% against V-PCC inter-frame.
• Performance gains were cross-checked and verified within the MPEG working group.

Conclusions:

• The proposed deep learning-based inter-frame encoding scheme offers substantial improvements in dynamic point cloud geometry compression efficiency.
• This method provides a promising direction for next-generation point cloud compression standards, enabling more efficient data handling for real-time and high-fidelity applications.