GRNet:Geometry Restoration for G-PCC Compressed Point Clouds Using Auxiliary Density Signaling.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Data Compression

Background:

• Lossy Geometry-based Point Cloud Compression (G-PCC) degrades point cloud geometry, impacting reconstruction quality and downstream tasks like classification.
• G-PCC distortions stem from point vanishing and point displacement, influenced by original point cloud density.

Purpose of the Study:

• To propose GRNet, a novel network for geometry restoration of large-scale G-PCC compressed point clouds.
• To address the dual challenges of point vanishing and displacement caused by G-PCC.

Main Methods:

• GRNet employs two specialized models: Coordinate Expansion for point vanishing and Coordinate Refinement for point displacement.
• A Feature Analysis Module with kNN-based Transformer fuses multiscale information for adaptive neighborhood geometric characterization.
• Auxiliary density information guides the selection of restoration strategies.

Main Results:

• GRNet significantly improves G-PCC anchor performance.
• The method demonstrates superior quantitative and qualitative results across diverse point cloud types (solid, dense, sparse).
• GRNet achieves faster processing speeds and utilizes smaller models compared to existing learning-based methods.

Conclusions:

• GRNet effectively restores geometry in G-PCC compressed point clouds, mitigating quality degradation.
• The proposed approach offers an efficient and effective solution for point cloud geometry restoration, suitable for practical applications.