Spectrum Alignment for Robust 3D Point Cloud Correspondences Estimation.

Area of Science:

• Computer Vision
• Geometry Processing
• Computational Geometry

Background:

• Estimating dense point-to-point correspondences between 3D point clouds is crucial for applications like texture and motion transfer.
• Non-rigid transformations in shapes pose significant challenges for traditional correspondence methods.
• Existing methods often rely on spectral representations using Laplace-Beltrami Operator (LBO) eigenfunctions but suffer from misalignment and outliers.

Purpose of the Study:

• To develop an efficient approach for aligning spectra of LBOs in 3D point clouds.
• To enhance the robustness of function correspondence estimation against outliers.
• To improve the accuracy of dense point-to-point correspondences for approximately isometric shapes.

Main Methods:

• Spectral alignment of Laplace-Beltrami Operators for 3D point clouds.
• Robust function correspondence estimation techniques to handle outliers.
• Comparison with state-of-the-art 3D shape-matching algorithms on benchmark datasets.

Main Results:

• Successfully aligned LBO spectra, maintaining compatible eigenfunctions for non-rigid shapes.
• Demonstrated improved robustness in function correspondence estimation.
• Achieved effective 3D shape matching, outperforming existing methods on benchmark datasets.

Conclusions:

• The proposed method offers an efficient and robust solution for dense point-to-point correspondences in challenging 3D shape matching scenarios.
• The spectral alignment and outlier rejection techniques are key to handling non-rigid transformations effectively.
• This work advances the field of geometry processing with practical implications for computer graphics and vision.