MATLAB: Computation of Match Metric in matchFeatures

computer visionMATLAB

I'm using the matchFeatures function to match two sets of SURF features. Running with the default arguments, I know that sum-of-squares distances (SSD) are computed and both a threshold test and ratio test are used. I'm a little curious at how the threshold test works. Ratio tests and nearest neighbor matching are well documented.

From generating C code output, I know with the default threshold parameter of 1%, match metrics greater than 0.4 are rejected. This is because (per the documentation) ideal matches have a metric of 0 and the worst matches have a metric of 4. Everything else is buried in library code.

If I were trying to replicate this threshold test in Python (or even C++ without Coder) and OpenCV, what kind of math would I be doing to compute the metrics? Why is the max distance 4? Is there some sort of normalization or statistical technique being used?

Best Answer

Hi,

Each feature vector is normalized before computing similarity score between feature vectors. This brings each value of feature vector to range [-1,1].

    Normalization =(fVec-mean(fVec))/norm(fVec);

SSD is computed using formula

    (|| fVec1 - fVec2 ||L2 ) / || fVec2||);

and match score computed between normalized feature vectors rather than exact feature vectors.

Ideal matches have a metric of 0 and the worst matches have a metric of 4 as maximum value of the above equation is 4 when fVec1 = -fVec2 and min is 0 when fVec1 = fVec2.

Hope it helps!!

Related Solutions

MATLAB: MatchFeatures returns wrong matches

Hi Manos,

matchFeatures is not really meant to be used to match the (x,y) locations of points. It is designed to match feature descriptors, like SURF, BRISK, or FREAK. You are seeing these strange results, because matchFeatures interprets its inputs as feature vectors, and normalizes them to have a magnitude of 1.

For correct usage of matchFeatures please see the following example.

MATLAB: MatchFeatures gives two different answers

They are not expected to share the same length. The first tries to find the best matches for features1 in features2 and the second tries to find the best matches for features2 in features1.

For the nearest neighbor ratio matching method (the default), this plays out as follows:

Assume features1 has 100 features and features2 has 200.

In the case of matchFeatures(features1,features2), for each of the 100 features in features1, the top 2 matches (2 nearest neighbors) from the 200 features in features2 are found. This gives us 100 pairs of 2 features each. These are then accepted or rejected by seeing whether the distance ratio between the nearest neighbor and the second nearest neighbor.

The process I described above is not symmetric. So, you can expect that matchFeatures(features2,features1) maybe slightly different.

You can look at this paper for details about this: K. Mikolajczyk and C. Shmid, "A Performance Evaluation of Local Descriptors," IEEE PAMI, 27, 10 (2005)

If you would like symmetric matches, try the 'NearestNeighborSymmetric' method:

    indexPairs1 = matchFeatures(features1,features2,'Method','NearestNeighborSymmetric');
    indexPairs2 = matchFeatures(features2,features1,'Method','NearestNeighborSymmetric');
>> isequal(fliplr(sort(indexPairs2)),sort(indexPairs1))
ans =
       1

Best Answer

Related Solutions

MATLAB: MatchFeatures returns wrong matches

MATLAB: MatchFeatures gives two different answers

Related Question