3-D Point Cloud Attribute Compression With -Laplacian Embedding Graph Dictionary Learning.

Area of Science:

• Computer Vision
• Data Compression
• Geometric Deep Learning

Background:

• 3-D point clouds offer rich scene information but pose significant compression challenges.
• Existing compression methods struggle to utilize irregular signal statistics and high-order geometric structures effectively.

Purpose of the Study:

• To develop an advanced 3-D point cloud attribute compression framework.
• To address limitations of current methods in exploiting complex geometric data structures.

Main Methods:

• Proposed a novel p-Laplacian embedding graph dictionary learning framework.
• Formulated a nonconvex minimization problem with p-Laplacian embedding regularization.
• Employed an alternating optimization paradigm using ADMM for efficient solution.

Main Results:

• Achieved superior M-term approximation and point cloud attribute compression performance.
• Outperformed state-of-the-art transform-based methods and MPEG G-PCC reference software.
• Demonstrated effective exploitation of varying signal statistics and high-order geometric structures.

Conclusions:

• The proposed p-Laplacian embedding graph dictionary learning framework is the first of its kind for point cloud compression.
• The integrated layered compression scheme efficiently exploits 3-D point cloud correlations.
• This approach offers a significant advancement in efficient 3-D point cloud data handling.