HA-TiNet: Learning a Distinctive and General 3D Local Descriptor for Point Cloud Registration.

Area of Science:

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

Background:

• Extracting geometric features from 3D point clouds is crucial for tasks like registration and recognition.
• Existing methods often struggle with rotation invariance and generalization.

Purpose of the Study:

• To develop a distinctive, general, and rotation-invariant 3D local descriptor.
• To improve the performance of 3D point cloud analysis tasks.

Main Methods:

• Proposed HA-TiNet (height-azimuth image based transformation-invariant net) method.
• Utilizes a height-azimuth image generator based on a local reference axis (LRA).
• Employs a ResNet-based backbone with a rotation-invariant layer for feature extraction.

Main Results:

• HA-TiNet generates rotation-invariant height-azimuth images with high accuracy.
• The method achieves full rotation-invariance by removing z-axis rotation variance.
• Demonstrated superior performance, descriptiveness, and generalization on indoor and outdoor datasets.

Conclusions:

• HA-TiNet offers a simple yet effective solution for 3D local descriptor learning.
• The approach significantly outperforms state-of-the-art descriptors.
• The method shows strong potential for various 3D point cloud applications.