LRF-Net: Learning Local Reference Frames for 3D Local Shape Description and Matching

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Local Reference Frames (LRFs) are crucial for 3D shape analysis and matching.
• Existing hand-crafted LRFs often lack repeatability and robustness.
• There is a need for more adaptive and reliable LRF methods.

Purpose of the Study:

• To develop a novel, data-driven approach for learning Local Reference Frames (LRFs).
• To improve the repeatability and robustness of LRFs in 3D shape analysis.
• To enhance performance in 3D point cloud matching and pose estimation tasks.

Main Methods:

• A Siamese network architecture was employed for weakly supervised LRF learning.
• Learned weights were used to measure the contribution of neighboring points to LRF construction.
• The proposed method, LRF-Net, was trained and evaluated on public 3D datasets.

Main Results:

• LRF-Net significantly outperforms state-of-the-art LRF methods in repeatability and robustness.
• Achieved 0.686 MeanCos performance on the UWA 3D modeling dataset, surpassing the closest method by 0.18.
• Demonstrated improved local shape description and 6-DoF pose estimation accuracy.

Conclusions:

• The proposed LRF-Net offers a robust and repeatable solution for LRF construction.
• Learned LRFs provide a significant advantage over hand-crafted methods in 3D computer vision tasks.
• LRF-Net effectively enhances 3D point cloud matching and pose estimation capabilities.