UGD: An Unsupervised Geometric Distance for Evaluating Real-World Noisy Point Cloud Denoising.

Area of Science:

• Computer Vision
• Geometric Processing
• Machine Learning

Background:

• Point cloud denoising is vital for real-world applications.
• Current evaluation metrics require ground-truth data, which is often unavailable.
• This limitation hinders the assessment of denoising methods in practical scenarios.

Purpose of the Study:

• To propose a novel unsupervised geometric distance (UGD) for evaluating point cloud denoising.
• To enable quantitative assessment using only noisy point clouds.
• To overcome the limitations of supervised metrics in real-world applications.

Main Methods:

• A patch-wise prior model is learned from clean point clouds using a Gaussian Mixture Model (GMM).
• This GMM serves as a ground truth to measure geometric variations in denoised point clouds.
• A self-supervised learning framework with multi-task learning trains the feature extraction network.

Main Results:

• The proposed UGD achieves performance comparable to supervised full-reference metrics on synthetic data.
• Experiments on real-world data confirm UGD's effectiveness for unsupervised evaluation.
• The method successfully evaluates denoising methods using only noisy point clouds.

Conclusions:

• UGD provides a robust and practical solution for evaluating point cloud denoising.
• It eliminates the need for ground-truth data, making it suitable for real-world scenarios.
• This unsupervised metric advances the field of point cloud processing and analysis.