Simultaneous nonrigid registration of multiple point sets and atlas construction

Area of Science:

• Computer Vision
• Medical Imaging
• Computational Geometry

Background:

• Groupwise registration of unlabeled point sets is complex due to the need for point correspondence in nonrigid settings.
• Existing methods often struggle with establishing accurate correspondences for complex shape variations.

Purpose of the Study:

• To develop a robust algorithm for simultaneous groupwise registration and mean shape computation of unlabeled point sets.
• To overcome the challenge of point correspondence in nonrigid registration tasks.

Main Methods:

• A novel algorithm that computes a mean shape (probability density function) from multiple unlabeled point sets (finite-mixture models).
• Simultaneous nonrigid registration of point sets to the emerging mean shape by minimizing Jensen-Shannon (JS) divergence.
• Derivation of the analytic gradient of the JS divergence for efficient optimization.

Main Results:

• The algorithm successfully performs groupwise registration without explicit point correspondence.
• It generates a probabilistic atlas as a by-product, representing the aligned shapes.
• Experimental results demonstrate effectiveness on 2D and 3D synthetic and real data.

Conclusions:

• The proposed JS divergence minimization method offers a robust solution for nonrigid groupwise registration of unlabeled point sets.
• This approach is valuable for applications like creating shape atlases and registering 3D range data in computer vision and graphics.
• The algorithm's ability to avoid correspondence problems makes it particularly useful for complex datasets.