DNet: Dynamic Neighborhood Feature Learning in Point Cloud.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Neighborhood selection is crucial for point cloud feature learning.
• Current methods often use fixed neighborhoods, potentially limiting performance.
• The rationality of neighborhood selection impacts point cloud processing outcomes.

Purpose of the Study:

• To propose a novel point cloud learning network, the Dynamic neighborhood Network (DNet).
• To enable dynamic neighborhood selection for more effective point feature learning.
• To enhance the performance of point cloud processing tasks.

Main Methods:

• Developed the Dynamic neighborhood Network (DNet) with a multi-head structure.
• Incorporated a Feature Enhancement Layer (FELayer) to enrich manifold features.
• Implemented a masking mechanism to filter low-contribution neighborhood points.

Main Results:

• DNet dynamically learns manifold and spatial geometric features.
• The masking mechanism identifies effective neighborhood relationships for each point.
• Experimental results show DNet's superiority over state-of-the-art methods on public datasets.

Conclusions:

• DNet offers a dynamic approach to neighborhood selection in point cloud learning.
• The proposed network effectively learns point features by considering adaptive neighborhoods.
• DNet demonstrates significant competitiveness in various point cloud processing applications.