PCDNF: Revisiting Learning-Based Point Cloud Denoising via Joint Normal Filtering.

Area of Science:

• Computer Vision
• Geometry Processing
• Machine Learning

Background:

• Point cloud denoising is crucial in 3D data processing but remains challenging.
• Current methods often treat denoising and normal filtering separately, limiting performance.
• The interdependence between point cloud noise and normal inaccuracies is often overlooked.

Purpose of the Study:

• To propose an end-to-end network for joint normal filtering and point cloud denoising.
• To improve noise removal accuracy while preserving fine geometric details.
• To leverage multitask learning for enhanced point cloud processing.

Main Methods:

• Developed PCDNF, an end-to-end network for joint normal filtering-based point cloud denoising.
• Introduced an auxiliary normal filtering task to improve noise reduction and feature preservation.
• Designed a shape-aware selector using latent tangent space representations.
• Implemented a feature refinement module to fuse point and normal features.

Main Results:

• The proposed PCDNF method outperforms state-of-the-art approaches in point cloud denoising.
• Achieved superior performance in normal filtering tasks compared to existing methods.
• Demonstrated effective preservation of geometric features, including sharp edges and corners.

Conclusions:

• Jointly addressing normal filtering and point cloud denoising is more effective than separate approaches.
• The novel modules in PCDNF significantly enhance noise removal and feature recovery.
• PCDNF offers a robust solution for high-quality point cloud processing.