Unsupervised Learning of Local Equivariant Descriptors for Point Clouds.

Area of Science:

• 3D Computer Vision
• Geometric Deep Learning
• Machine Learning

Background:

• 3D keypoint correspondences are crucial for 3D computer vision and graphics.
• Learned local descriptors surpass handcrafted ones but require extensive labeled data.
• Existing unsupervised methods for descriptor learning underperform supervised approaches.

Purpose of the Study:

• To develop an unsupervised method for learning 3D local descriptors that overcomes limitations of current supervised and unsupervised techniques.
• To improve generalization and reduce reliance on data augmentation for viewpoint invariance.
• To achieve competitive performance with supervised methods, especially in transfer learning.

Main Methods:

• Proposed Local Equivariant Descriptor (LEAD) learning an equivariant 3D local descriptor.
• Utilized Spherical Convolutional Neural Networks (CNNs) for equivariant representation learning.
• Employed plane-folding decoders for unsupervised learning.

Main Results:

• LEAD significantly outperforms existing unsupervised 3D local descriptor methods on standard surface registration datasets.
• LEAD achieves results competitive with supervised approaches, demonstrating strong transfer learning capabilities.
• The equivariant approach effectively addresses viewpoint invariance without compromising generalization.

Conclusions:

• Learning equivariant descriptors is a viable alternative to invariant ones, overcoming data requirements and generalization issues.
• LEAD offers a powerful unsupervised solution for 3D local descriptor learning.
• The method shows significant promise for real-world applications in 3D computer vision and graphics.