One Shot Learning for Edge Detection on Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• 3D scanners have unique data characteristics and sampling error distributions.
• Training neural networks on diverse scanner data yields suboptimal edge extraction performance.
• Scanner-specific data distributions necessitate specialized network training approaches.

Purpose of the Study:

• To develop a novel one-shot learning method for accurate edge extraction on 3D point clouds.
• To address the challenge of varying data distributions from different 3D scanners.
• To improve the performance of point cloud analysis by learning target-specific data distributions.

Main Methods:

• Proposed OSFENet (One-Shot Feature Extraction Network), a lightweight network for edge extraction.
• Designed a filtered-KNN-based surface patch representation for one-shot learning.
• Introduced an RBF_DoS module integrating Radial Basis Function-based Descriptor of the Surface patch.

Main Results:

• OSFENet achieved superior results compared to 7 baselines on the ABC dataset.
• Demonstrated effectiveness across diverse real-scanned datasets, including indoor (S3DIS) and outdoor (Semantic3D, UrbanBIS) scenes.
• Validated the practical utility of the proposed one-shot learning approach for edge extraction.

Conclusions:

• The proposed one-shot learning method effectively learns scanner-specific data distributions for improved edge extraction.
• OSFENet offers a robust and efficient solution for point cloud edge extraction in varied real-world scenarios.
• The approach significantly advances the field of 3D point cloud analysis by enabling high-performance, scanner-adaptive feature extraction.