Toward the Repeatability and Robustness of the Local Reference Frame for 3D Shape Matching: An Evaluation

Area of Science:

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

Background:

• Local Reference Frames (LRFs) are crucial for 3D local feature descriptors, offering rotational invariance and spatial information.
• Existing LRF methods lack comprehensive comparative analysis regarding performance under various conditions.

Purpose of the Study:

• To conduct a thorough evaluation of eight state-of-the-art LRF methods.
• To assess LRF repeatability and robustness across diverse 3D data modalities and application scenarios.
• To identify the strengths and weaknesses of current LRF techniques.

Main Methods:

• Experimental evaluation of eight LRF methods on six benchmarks.
• Testing across different data modalities (LiDAR, Kinect, Space Time) and applications (shape retrieval, 3D registration, object recognition).
• Assessing robustness to nuisances like noise, varying resolution, occlusion, and keypoint errors; also evaluating computational efficiency.

Main Results:

• Comparative performance analysis of LRFs under various conditions and data types.
• Identification of LRFs most suitable for specific applications and robustness requirements.
• Detailed assessment of how different LRFs interact with keypoint detectors and feature representations.

Conclusions:

• Summary of the traits, advantages, and limitations of the evaluated LRF methods.
• Provides guidance for selecting appropriate LRFs based on application needs and data characteristics.
• Highlights areas for future research in LRF development for 3D computer vision.