Robustness of RANSAC

[lustiz]

Hi there, question: How exactly robust is RANSAC supposed to be? There is a couple of parameters for the stitching process, i.e. number of iterations (depending on our probability estimate of samples being inliers), the chosen threshold for small distances as well as the decision whether we accept a sample together with its computed homography or not.

So even if I
1) choose a pretty conservative guess of the inlier probablity (-> a rather big number of iterations),
2) throw in a couple of different thresholds (doesn't matter that much) and
3) do only accept homographies up to a certain condition number (because samples may be ill-posed)

the resulting stitch is not always as appealing as with the given solution in 'H.mat'. Is there a certain trick to always get good results? Or can we expect RANSAC fail to deliver a good solution every now and then?

Thanks!

[lustiz]

OK, got it pretty robust, now. The trick was to choose a pretty conservative guess for the inlier probability (so that you have a bunch of iterations, though still below 100) together with a higher threshold for the chi-square-distances. You do not have to worry about a rather strict threshold, RANSAC will handle the rest


BTW, a tip for visual debugging: I found it quite nice to draw different lines for different correspondences so that distinguishing is more easy...
You can do that by defining a colormap for the correspondence lines...

Code: Alles auswählen

 
% Different colors for different lines
cmap = hsv(numPairs);
% random index s.t. neighboring lines may have different colors
ridx = randperm(numPairs);
for i = 1:size(pairs, 1)
  currPoints = ...
  plot(currPoints, 'color', cmap(ridx(i), :));
end