MPED: Quantifying Point Cloud Distortion Based on Multiscale Potential Energy Discrepancy.

Area of Science:

• Computer Vision
• 3D Data Processing
• Geometric Deep Learning

Background:

• Effective distortion quantification for point clouds is crucial for both human and machine perception tasks.
• Current methods fail to meet the requirements of being differentiable, distortion discriminable, and computationally efficient.
• A gap exists in methods that can accurately predict subjective human scores and serve as loss functions for deep learning.

Purpose of the Study:

• To propose a new distortion quantification method for point clouds called multiscale potential energy discrepancy (MPED).
• To develop a method that is differentiable, distortion discriminable, and computationally efficient.
• To provide a superior alternative to existing distortion quantification techniques for diverse point cloud applications.

Main Methods:

• Introduced a novel point cloud feature description method, point potential energy (PPE), inspired by classical physics.
• Developed MPED by evaluating potential energy across various neighborhood sizes to achieve multiscale distortion capture.
• Demonstrated that Chamfer distance is a special case of the proposed MPED.

Main Results:

• MPED effectively balances global and local information, capturing distortion in a multiscale manner.
• Extensive experiments confirmed MPED's superiority over current methods in both human and machine perception tasks.
• The proposed method addresses the limitations of existing distortion quantification techniques.

Conclusions:

• MPED offers a robust and versatile solution for point cloud distortion quantification.
• The method shows significant improvements for tasks including compression, enhancement, reconstruction, completion, and upsampling.
• MPED provides a valuable tool for advancing research and applications in 3D point cloud processing.