OpenCV - SURF Feature Comparison

Question

I am curious about how the OpenCV Feature descriptors are compared. For instance, I can use cvExtractSURF() to get a list of features and their 64-bit (or 128-bit) descriptors, where can I find out how two descriptor can be compared?

In stepping through some sample code, to me it looks like two of my "matched" features have very different descriptors (at least by numerical values).

Has anyone ever figures out how to take two descriptor arrays and compare them?

Googling hasn't helped too much...

Cheers, Brett

Answer 1

You might want to look at the paper Local invariant feature detectors: a survey. It's a great paper with a description of widely used feature detectors, including SURF.

Answer 2

In the OpenCV 2.1 sample file find_obj.cpp, two methods are presented:

• the built-in C++ Flann function(flann give aproximate solution and works faster), I don't know exactly how it works, but it is documented here.
• a simpler C function (findPairs()) which finds the nearest neighbor by computing a simple euclidian distance between descriptors (look at the compareSURFDescriptors() function). The laplacian may also be used as a first indicator of similarity, as matching points have not the same laplacian (1 or -1). This sample is available here.