Evaluating Local Geometric Feature Representations for 3D Rigid Data Matching.

Area of Science:

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

Background:

• Local geometric descriptors are crucial for 3D rigid data matching.
• Existing evaluations primarily focus on local reference frames (LRFs) or overall descriptors, neglecting feature representation comparisons.
• A quantitative evaluation of feature representations is needed.

Purpose of the Study:

• To comprehensively evaluate and compare nine state-of-the-art local geometric feature representations.
• To provide a convincing ranking of feature representations by assessing them against ground-truth LRFs.
• To offer new perspectives on local geometric feature description.

Main Methods:

• Evaluation of nine feature representations using ground-truth LRFs.
• Experiments conducted on six standard datasets across diverse scenarios (shape retrieval, registration, recognition) and modalities (LiDAR, Kinect, Space Time).
• Assessment under various perturbations: Gaussian noise, shot noise, data decimation, clutter, occlusion, and limited overlap.

Main Results:

• Comprehensive analysis of feature representations based on distinctiveness, robustness, compactness, and efficiency.
• Experimental outcomes reveal significant findings that challenge existing assumptions.
• A comparative summary of evaluated methods is provided to guide practical applications.

Conclusions:

• The study provides a thorough quantitative comparison of local geometric feature representations.
• Findings offer valuable insights for selecting and developing descriptors for 3D data matching tasks.
• This work serves as a crucial reference for researchers and practitioners in the field.