APN-Net: An Adaptive Perception Network for Point Cloud Normal Estimation.

Area of Science:

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

Background:

• Surface normal estimation is vital for point cloud processing and downstream applications.
• Current methods struggle with scale ambiguity and density variations in point clouds.
• Unstructured point clouds present challenges for traditional feature extraction techniques.

Purpose of the Study:

• To propose APN-Net, an adaptive perception network for robust point cloud normal estimation.
• To address the limitations of existing methods in handling scale ambiguity and varying point cloud densities.
• To improve the accuracy and efficiency of normal estimation in complex 3D environments.

Main Methods:

• Developed the Graphical Information Self-perception (GIS) module for implicit region partitioning and expanded receptive fields.
• Introduced the Adaptive Graph Convolution (AGC) module with adaptive kernels for richer feature representation.
• Utilized both synthetic and real-world scan datasets for comprehensive evaluation.

Main Results:

• APN-Net effectively extracts both local geometric details and global structural information.
• The proposed GIS and AGC modules alleviate scale ambiguity and capture complex geometric features.
• Achieved superior performance in unoriented normal estimation, especially for point clouds with significant density variations.

Conclusions:

• APN-Net offers a significant advancement in point cloud normal estimation.
• The adaptive perception approach provides robustness against scale ambiguity and density variations.
• Demonstrated superior performance compared to existing methods on diverse datasets.