WSDesc: Weakly Supervised 3D Local Descriptor Learning for Point Cloud Registration.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Point Cloud Processing

Background:

• 3D point cloud registration is crucial for many applications.
• Existing methods often rely on hand-crafted features or fully supervised learning.
• Learning robust local descriptors from 3D data remains a challenge.

Purpose of the Study:

• To develop a novel method for learning 3D local descriptors in a weakly supervised manner.
• To improve the robustness and accuracy of point cloud registration.
• To address limitations of fixed-size local support in descriptor extraction.

Main Methods:

• Proposed WSDesc, a weakly supervised descriptor learning method.
• Introduced a differentiable voxelization layer for data-driven support size optimization.
• Developed a novel registration loss based on transformation rigidity and partial overlap priors.

Main Results:

• WSDesc learns adaptive local support for 3D descriptors.
• The method achieves superior performance on geometric registration benchmarks.
• Weak supervision eliminates the need for ground-truth alignment information.

Conclusions:

• WSDesc offers a robust and efficient approach to 3D local descriptor learning.
• The proposed method advances weakly supervised learning for point cloud registration.
• Learned descriptors generalize well and improve registration accuracy.