You Only Train Once: Learning General and Distinctive 3D Local Descriptors.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Geometry

Background:

• Extracting robust 3D local features is crucial for tasks like point cloud registration.
• Existing methods often use noise-sensitive handcrafted features or rotation-variant neural networks, limiting their generalizability.
• Learning general and robust local feature descriptors for surface matching remains a challenge.

Purpose of the Study:

• To propose SpinNet, a novel neural network for extracting distinctive, robust, and general 3D local surface descriptors.
• To achieve rotation-invariant feature descriptors suitable for surface matching.
• To improve the generalization ability of local feature descriptors across diverse scenarios.

Main Methods:

• Introduced a Spatial Point Transformer to create a rotation-equivariant cylindrical representation of local surfaces.
• Developed a Neural Feature Extractor with point-based and 3D cylindrical convolutional layers for geometric pattern learning.
• Utilized an invariant layer to generate rotation-invariant feature descriptors.

Main Results:

• SpinNet demonstrated superior performance compared to state-of-the-art methods on both indoor and outdoor datasets.
• The proposed method achieved significant improvements in point cloud registration and surface matching tasks.
• SpinNet exhibited excellent generalization capabilities across unseen scenarios and different sensor modalities.

Conclusions:

• SpinNet provides a simple yet effective solution for learning robust and general 3D local feature descriptors.
• The rotation-invariant nature and strong generalization ability of SpinNet make it highly suitable for real-world applications.
• This work advances the field of 3D feature extraction for computer vision and robotics.