From Noise Addition to Denoising: A Self-Variation Capture Network for Point Cloud Optimization.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• 3D scanner data (point clouds) often contain noise, hindering their use in advanced applications.
• Existing denoising methods may struggle with uniform noise distribution and underlying surface drift.

Purpose of the Study:

• To propose a novel point cloud optimization method for denoising and homogenizing noisy point clouds.
• To develop a network that captures commonalities between noisy and self-variation point clouds for noise removal.

Main Methods:

• A Self-Variation Capture Network (SVCNet) is proposed, generating a self-variation point cloud through noise perturbation.
• SVCNet captures commonalities between the original and self-variation point clouds in the latent space via feature aggregation and averaging.
• An edge constraint module is incorporated to mitigate low-pass filtering effects during denoising.

Main Results:

• The proposed method effectively denoises point clouds without assuming specific noise characteristics.
• It successfully filters drift noise and achieves a uniform point cloud distribution.
• Experimental results demonstrate superior performance compared to state-of-the-art algorithms, particularly in point cloud uniformity.

Conclusions:

• The SVCNet method offers an effective approach for point cloud denoising and homogenization.
• The algorithm shows promise for applications beyond denoising, including point cloud upsampling.