3

Basically I have:

BruteForceMatcher<L2<float>>().knnMatch(descriptor1,descriptor2,matches,2);

To get only good matches I parse all the "matches" vectors and check the distance like this:

if( (matches[i][0].distance / matches[i][1].distance) < ratio ){
  //> Good match!
}

But what matches[i][0].distance mean ? The distance between matches[i][0] and ?

My supposition

For what I could guess it would sound more logic to me to calculate the euclian distance between the first match with it's NN, and filter it with a threshold, something like:

//> I calculate the distance between 2 nearest neighborhood and filter it based on thresold
foreach( matches : i) {
 if ( euclianDistance( matches[i][0] , matches[i][1] ) < threshold ) {
   //> good match
 }
}
  • 1
    The idea of this filtering is to remove unreliable matches - matches where we can not say which one is better (looking at descriptors only). And the original rule is better then your suggestion for this task: imagine a circle - the center of the circle - is our train descriptor. And two best matches - are the points lying on the opposite ends of the diameter. In this case the distance between these matches will be big but they are indistinguishable from the point of train descriptor. – Andrey Kamaev Jun 24 '12 at 10:30
  • @Andrey: Just one thing: how 2 best matches can be best if they lie on the oppiste ends of the diameter? I mean BruteForceMatcher compares each train descriptors with each of query descriptors calculating their euclian distance (or whatever...) Maybe i just don't get what `.distance` really means. (I am upvoting anything you write anyway :) –  Jun 24 '12 at 10:36
  • 1
    You are correct, BruteforceMatcher is doing exhaustive search - it compares each descriptor from train with each descriptor from query. And `knnMatch` is simply returning 2 (N) descriptors with smallest distances. And there are nothing unusual if 2 best having the same distance. My previous comment illustrates how it can look like in 2D case - two closest are equidistant and others are outside that circle (knnMatch will not return them for N=2) – Andrey Kamaev Jun 24 '12 at 10:46

1 Answers1

4

descriptor - is a point of N-dimensional space.

match - is a pair of descriptors - one from the first set and one from the second set (also called train and query sets).

distance - is a L2 metric for 2 descriptors pointed by the match structure. (You are specifying the type of metric as a template parameter for BruteForceMatcher).

match[i][0].distance = L2(descriptor1.row(match[i][0].trainIdx),
                          descriptor2.row(match[i][0].queryIdx))

So knnMatch returns two closest descriptors from the query set for every descriptor from the train set. Next you are filtering out cases when the two found descriptors are close to each other.

Andrey Kamaev
  • 29,582
  • 6
  • 94
  • 88
  • It might be the same keypoint. But if the images has any difference then the descriptors of the same keypoint will be also different and with good probability descriptor from another keypoint can be closer. – Andrey Kamaev Jun 24 '12 at 10:23
  • can you tell me what distance is calculated inside the kmmMatch method. Is it Euclidean distance or what. Thanks. – posha Nov 06 '13 at 16:53