Matching algorithm coding

This week end, I coded a simple version of the matching algorithm. I am using OpenCV library, and defined some C++ classes: minutiaPoint, minutiaePoints. I coded the function that find the optimal affine transformation between N correspondences (that is mainly used for N=3, that gives an exact solution).

I coded the ransac algorithm using the transformation function. It needs 2 sets of minutiae points, each minutia in the first set has some correspondences in the other set. Then 3 points are chosen randomly in the first set, and their correspondences are chosen randomly in the set of possible correspondences.

To see the results, I used the minutiae points posted last time. The red crosses correspond to the first set of points, the green ones are for the second set. The first image is the unmatched version (surimposed points).




Then find and apply the transformation using 3 close points (in green in the bottom of the image). This gives a first alignment, which is not so good (especially with far points).




Then I used 3 far points to estimate the transformation (one at the top, one on the left, one at the bottom). The alignment is better.




The best alignment found is by using all the 27 minutiae points and estimating the transformation. This shows that in ransac, it is good to reestimate the transformation using all the inliers found (the points that are found to match). It is not good to estimate the transformation using all the points as there are a lot of outliers that will influence the results. That's the job of the ransac algorithm: get rid of the outliers.