Fast and Robust Non-Rigid Registration Using Accelerated Majorization-Minimization.

Area of Science:

• Computer Vision
• Medical Imaging
• Computational Geometry

Background:

• Non-rigid 3D registration aligns source and target 3D shapes through deformation.
• Challenges include noisy data, outliers, and partial overlaps, leading to slow convergence in existing methods.
• Current methods often use l_p norms and proximal algorithms, limiting practical application.

Purpose of the Study:

• To develop a robust non-rigid registration formulation that effectively handles imperfect data.
• To improve the convergence speed and efficiency of registration algorithms.
• To outperform existing methods in accuracy and computational speed for 3D shape alignment.

Main Methods:

• Proposed a novel formulation for robust non-rigid registration using a globally smooth robust norm.
• Employed the majorization-minimization algorithm for optimization, reducing iterations to convex quadratic problems.
• Integrated Anderson acceleration to enhance solver convergence speed for efficient computation.

Main Results:

• The proposed method demonstrates effectiveness in non-rigid alignment of 3D shapes with outliers and partial overlaps.
• Quantitative evaluations show superior registration accuracy compared to state-of-the-art methods.
• Achieved significant improvements in computational speed, enabling efficient processing on limited hardware.

Conclusions:

• The developed robust non-rigid registration method offers a significant advancement in handling challenging 3D data.
• The combination of a globally smooth norm and accelerated optimization provides a fast and accurate solution.
• The method is well-suited for applications requiring efficient and precise 3D shape alignment, even with noisy or incomplete data.