Non-Rigid Point Matching
Point Matching: Problem & Algorithm
I have been primarily interested in the problem of solving for high dimensional non-rigid deformations given two sets of feature points, for which the correspondence information is unknown beforehand. This is what we call the "non-rigid point matching problem."
| We have
designed a new non-rigid point matching algorithm that is capable of
estimating both complex non-rigid transformations as well as meaningful
correspondences between two sets of points. The effectiveness of the
algorithm comes from two techniques: softassign and deterministic
The algorithm is very robust. First, it tolerates noises. Second, it can automatically evalute all evidence and reject outliers. Finally, it demonstrates stong ability in overcoming local minima and bad initializations. The algorithm is hence called the "robust point matching algorithm (RPM)."
A demo for robust point matching.
| This package
of MATLAB M-files provides a demo for the Robust Point Matching (RPM) algorithm.
Five example data point-sets are included. We also provide a simple GUI
to load the data and start the demo. All the source code (M-files) required
to execute RPM are included.
The code is provided under the terms of the GNU General Public License with an explicit clause permitting the M-files to be executed from within the Matlab environment.
Some references for robust point matching:
"A new algorithm for non-rigid point matching",
H. Chui and A. Rangarajan,
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2000, volume 2, pages 44-51.
Download the postscript file.
Honorable mention for Best Student Paper Award at CVPR 2000.
(2) "A feature
registration framework using mixture models",
(3) "A unified
framework for brain anatomical feature registration",
of cortical anatomical structures via 3D robust point matching",
(5) "A robust
point matching algorithm for autoradiograph alignment",
Further references on robust point matching can be found at Anand Rangarajan's homepage.
Acknowledgement: This work is partially supported by the National Science Foundation Grant IIS-9906081.