[Matching] Experiment with RANSAC test

[azasypkin]

In the scope of this bug I'd like to check if RANSAC test of matches can give us better score.

Blocks #6

[Yoric]

Out of curiosity, what are these tests, exactly?

[azasypkin]

Sure, here is the current matching flow:

• Extract keypoints and their descriptors from the images (train and query);
• Use BFMatcher to find matches;
• Apply ratio test, see description here. There is also SIFT whitepaper where ratio test is explained (section 7.1);

Cross-check test (short note about it here) is alternative to ratio test, but it performed worse for my test set (wrong matches for some objects got much bigger score). So I gave up on this for now.

What can be improved here is that after ratio test we can adjust score using other keypoints characteristics. E.g we have just a few matches and most of them are good ones (determined by ratio test), but due to their small number (let's say package is partially visible or we have glares) match isn't successful (score is below default 10 threshold). In this case we can try to bump up score value if we know that all these good matches have the same mutual arrangement as on the source picture. I don't know much about RANSAC test yet, but I'm going to figure out if it can give us what we want.