Norest-Net: Normal Estimation Neural Network for 3-D Noisy Point Clouds.

Area of Science:

• Computer Vision
• 3-D Data Processing
• Machine Learning

Background:

• 3-D point clouds from sensors like LiDAR and depth cameras are often noisy.
• Accurate normal estimation is vital for downstream 3-D processing tasks.
• Existing methods struggle to balance noise reduction and feature preservation.

Purpose of the Study:

• To develop a robust normal estimation method for noisy 3-D point clouds.
• To improve both noise filtering and surface feature preservation simultaneously.
• To provide a modular component for enhancing existing normal estimation techniques.

Main Methods:

• Proposed Norest-Net, a neural network with two specialized branches: NF-Net for noise filtering and NR-Net for feature refinement.
• NF-Net learns to predict ground-truth normals from noisy height map descriptors.
• NR-Net learns to predict ground-truth normals from bilateral-filtered point normal descriptors.

Main Results:

• Norest-Net significantly outperforms state-of-the-art methods in normal estimation accuracy.
• The method demonstrates superior performance in preserving surface features.
• Norest-Net exhibits enhanced robustness against noise in both synthetic and real-world data.

Conclusions:

• Norest-Net effectively addresses the trade-off between noise filtering and feature preservation in normal estimation.
• The proposed architecture offers a specialized and improved approach to 3-D point cloud processing.
• The detachable NR-Net module can enhance the performance of existing normal estimation algorithms.