Effective way to match SIFT descriptor

Question

There are 2 images A and B. I extract key points from them (a [i] and b [i]).
I wonder how to effectively determine the correspondence between [i] and b [j]?

The obvious method is to compare each point in with each point in B. But it takes a long time for large image databases. How can I just compare point a [i] only with b [k], where k has a small range?

I heard that kd-tree might be a good choice, right? Are there any good examples about kd-tree?

Any other suggestions?

Answer 1

The KD tree stores the processed descriptors in such a way that it is actually faster to find the most similar descriptor when doing a match.

It is very easy to use kd-tree in OpenCV, I will give you an example for a flaneograph:

flann::GenericIndex< cvflann::L2<int> > *tree; // the flann searching tree tree = new flann::GenericIndex< cvflann::L2<int> >(descriptors, cvflann::KDTreeIndexParams(4)); // a 4 kd tree 

Then, when you do the mapping:

 const cvflann::SearchParams params(32); tree.knnSearch(queryDescriptors, indices, dists, 2, cvflann::SearchParams(8)); 

Answer 2

The question is, do you really want to define a keyword between two images or calculate a similarity score .

If you want to determine the match, then I'm afraid you will have to look for brute force through all possible pairs of descriptors between the two images (there are several more advanced methods, such as FLANN - quick approximate nearest neighbor search, but acceleration is not significant if you have less or less than 2000 key points per image - at least in my experience). To get a more accurate match (not faster, just a better match), I can suggest you take a look at:

• DG Lowe. Distinctive features of the image from non-invariant key points - comparison with the second closest match
• J. Sivich and A. Zisserman. Google Video: a text search approach to an object matching in a video - spatial consistency section

If, on the other hand, you only need a measure of similarity in a large database, then a suitable place to start is:

• D. Nister and H. Stunius. Scalable recognition using the vocabulary tree - where they use a hierarchical approach based on a structure called the vocabulary tree to be able to calculate the similarity measure between the query image and the image from a large database.
• J. Sivich and A. Zisserman. Google Video: the text search approach to object matching in video is the same document as above, but it’s very useful to reinforce the approach in Nister, Stunius

Answer 3

OpenCV implements several strategies for mapping sets of key points. See the documentation for General descriptor interfaces>.