Nonrigid Point Set Registration With Robust Transformation Learning Under Manifold Regularization.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Nonrigid point set registration is crucial for tasks like object recognition and medical imaging.
• Existing methods struggle with noisy data and outliers, limiting their real-world applicability.

Purpose of the Study:

• To develop a robust transformation learning scheme for nonrigid point set registration.
• To improve accuracy and efficiency in learning transformations from noisy point sets.

Main Methods:

• Iterative establishment of point correspondences using local feature descriptors.
• Semisupervised learning framework with indicator variables to handle outliers in a mixture model.
• Manifold regularization to exploit intrinsic point set structure and a reproducing kernel Hilbert space model with sparsity-induced approximation for efficiency.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art techniques.
• Effective handling of degenerated data and noisy correspondences was observed.
• Successful application in learning motion flows for visual homing in mobile robot navigation.

Conclusions:

• The developed semisupervised learning approach offers a robust solution for nonrigid point set registration.
• The method's ability to manage outliers and leverage intrinsic data structure enhances registration accuracy.
• This work contributes a significant advancement for visual homing and related robotic navigation tasks.