MSL-Net: Sharp Feature Detection Network for 3D Point Clouds.

Area of Science:

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

Background:

• Sharp features are crucial geometric elements in 3D point clouds for tasks like reconstruction and registration.
• Existing detection methods struggle with noisy and non-uniformly dense point cloud data.

Purpose of the Study:

• To develop a robust and accurate deep-learning-based method for detecting sharp features in 3D point clouds.
• To overcome the limitations of current methods concerning data quality and density variations.

Main Methods:

• Proposed the Multi-scale Laplace Network (MSL-Net), utilizing an intrinsic neighbor shape descriptor.
• Established a discrete intrinsic neighborhood using a Laplacian graph to minimize surface estimation errors.
• Designed an intrinsic shape descriptor incorporating enhanced normal extraction and a cosine-based field estimation function.

Main Results:

• MSL-Net demonstrates a simple architecture capable of accurate feature prediction adhering to manifold distribution.
• The multi-scale structure provides strong analytical ability for local point cloud perturbations.
• Extensive experiments show MSL-Net outperforms state-of-the-art methods in robustness and accuracy.

Conclusions:

• MSL-Net offers a significant advancement in sharp feature detection for 3D point clouds.
• The method is robust to noise and density variations, outperforming existing approaches.
• MSL-Net provides accurate feature prediction with a simplified architecture and avoids complex calculations.