Deep Frequency Awareness Functional Maps for Robust Shape Matching.

Area of Science:

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

Background:

• Traditional functional maps struggle with frequency information in complex 3D shape matching.
• Significant deformations and topological inconsistencies degrade performance in existing methods.

Purpose of the Study:

• To introduce a novel unsupervised learning framework, Deep Frequency Awareness Functional Maps (DFAFM), for robust 3D shape matching.
• To enhance the adaptive capture of critical frequency information for improved functional map estimation.

Main Methods:

• Developed the Spectral Filter Operator Preservation constraint to ensure frequency information integrity.
• Learned spectral filters using orthonormal Jacobi polynomials with learnable coefficients.
• Incorporated filters as a loss function for joint supervision of functional maps, pointwise maps, and filters.
• Implemented a refinement strategy using learned filters to boost pointwise map accuracy.

Main Results:

• DFAFM significantly outperforms state-of-the-art methods on benchmark datasets.
• The framework demonstrates superior performance in challenging scenarios with non-isometric deformations and inconsistent topology.
• Achieved enhanced accuracy in 3D shape matching tasks.

Conclusions:

• DFAFM provides a robust and adaptive solution for 3D shape matching.
• The proposed frequency awareness approach overcomes limitations of traditional methods.
• The framework offers improved accuracy and generalization for diverse shape-matching problems.