PFF-Net: Patch Feature Fitting for Point Cloud Normal Estimation.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Data Processing

Background:

• Accurate normal estimation is crucial for 3D point cloud analysis.
• Existing methods struggle with varying data geometries and neighborhood size selection.
• Parameter-heavy strategies often lack efficiency and accuracy.

Purpose of the Study:

• To develop a robust and efficient method for normal estimation in point clouds.
• To address the challenge of selecting appropriate neighborhood sizes for diverse geometries.
• To improve the accuracy and speed of normal prediction for point cloud data.

Main Methods:

• A novel feature extraction technique using the fusion of multi-scale features.
• Patch Feature Fitting (PFF) model based on multi-scale features.
• Multi-scale feature aggregation and cross-scale feature compensation modules.

Main Results:

• Achieved state-of-the-art performance on synthetic and real-world point cloud datasets.
• Demonstrated superior accuracy and efficiency compared to existing methods.
• Reduced network parameters and running time.

Conclusions:

• The proposed multi-scale feature fusion method enables scale adaptation for varying local patches.
• The method provides an optimal feature description for robust normal estimation.
• This approach offers a significant advancement in point cloud processing.