GeoDualCNN: Geometry-Supporting Dual Convolutional Neural Network for Noisy Point Clouds

Area of Science:

• Computer Vision
• Geometric Deep Learning

Background:

• Point clouds are essential 3D data representations.
• Accurate normal estimation and denoising are crucial for point cloud processing.
• Existing methods struggle with noise robustness and feature preservation.

Purpose of the Study:

• To develop a novel deep learning model for simultaneous point cloud normal estimation and denoising.
• To leverage geometric domain knowledge for improved performance.
• To enhance noise robustness and feature preservation in point cloud processing.

Main Methods:

• A geometry-supporting dual convolutional neural network (GeoDualCNN) is proposed.
• The method defines a homogeneous neighborhood (HoNe) to handle surface discontinuities.
• Two parallel branches utilize geometric descriptors (homogeneous height map and normal map) for noise removal and feature preservation.

Main Results:

• GeoDualCNN achieves state-of-the-art performance in point cloud denoising and normal estimation.
• The model demonstrates superior noise-robustness compared to existing methods.
• Effective preservation of surface features is achieved during the denoising process.

Conclusions:

• GeoDualCNN offers a significant advancement in point cloud processing by integrating geometric priors into a deep learning framework.
• The proposed architecture effectively addresses the challenges of noise and feature preservation.
• This approach paves the way for more robust and accurate 3D data analysis.