Contrastive Learning for Joint Normal Estimation and Point Cloud Filtering.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Point cloud filtering and normal estimation are crucial in 3D data processing.
• Current methods often struggle with noise and preserving sharp geometric features like edges and corners.
• Separate processing of filtering and normal estimation limits performance.

Purpose of the Study:

• To develop a novel deep learning method for simultaneously estimating point cloud normals and filtering.
• To improve robustness to noise and enhance preservation of sharp geometric features.
• To outperform existing state-of-the-art methods in both tasks.

Main Methods:

• A 3D patch-based contrastive learning framework with noise corruption augmentation was used to train a robust feature encoder.
• The learned representations were fed into a regression network.
• A novel joint loss function was employed for simultaneous normal estimation and patch center displacement for filtering.

Main Results:

• The proposed method effectively performs joint normal estimation and point cloud filtering.
• Sharp features, fine details, and geometric integrity are well-preserved.
• Experimental results demonstrate superior performance compared to state-of-the-art techniques on both tasks.

Conclusions:

• The novel deep learning approach successfully integrates point cloud filtering and normal estimation.
• The method offers enhanced robustness to noise and superior preservation of geometric details.
• This work advances the capabilities of 3D data processing techniques.