Learning spectral descriptors for deformable shape correspondence.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Computational Geometry

Background:

• Feature descriptors are crucial for deformable shape analysis tasks like correspondence and registration.
• Spectral descriptors, such as Heat Kernel Signature (HKS) and Wave Kernel Signature (WKS), derived from the Laplace-Beltrami operator, are effective but can be further optimized.
• Current descriptors often lack task-specific optimization based on data statistics.

Purpose of the Study:

• To introduce a generic family of parametric spectral descriptors for deformable shape analysis.
• To develop a learning scheme for optimizing these descriptors by considering shape data statistics ('signal') and transformation invariances ('noise').
• To demonstrate the superiority of learned descriptors in shape correspondence tasks.

Main Methods:

• Formulated a generic family of parametric spectral descriptors.
• Developed a learning scheme inspired by the Wiener filter to optimize descriptors based on corpus statistics.
• Related the learning scheme to Mahalanobis metric learning.
• Evaluated descriptor performance on synthetic and scanned human figures for correspondence.

Main Results:

• The proposed learning scheme successfully constructs optimized spectral descriptors.
• The learned descriptors demonstrate superior performance in generating shape correspondences compared to existing methods.
• The approach shows robustness, enabling descriptors trained on synthetic data to transfer effectively to real scanned shapes.

Conclusions:

• Learned spectral descriptors offer a powerful and adaptable approach to deformable shape analysis.
• The proposed method provides a principled way to optimize descriptors for specific tasks and datasets.
• This work advances the state-of-the-art in shape analysis by enabling more accurate and robust shape correspondence.