BFMatcher (Bug #3172)

Using BFMatcher with "NORM_HAMMING" on binary descriptors (created by FREAK) gives unexpected results in certain conditions:

std::vector<Mat> descriptors; //FREAK descriptors of e.g. 1000 images
... //compute actual descriptors
BFMatcher bfMatcher(NORM_HAMMING, false);
bfMatcher.add(descriptors);
int nn = 10; //Number of nearest neighbours
bfMatcher.knnMatch(query_descriptors, matches, nn);

The above code returns invalid results, if nn happens to be smaller than any image's descriptor set.
For example one single image test.jpg might have only 9 (<nn) according FREAK descriptors.
The BFMatcher will calculate all distances using batchDistance (https://github.com/Itseez/opencv/blob/master/modules/features2d/src/matchers.cpp#L366).
Eventually it passes the current best results (nidx) to cv::batchDistance in https://github.com/Itseez/opencv/blob/master/modules/core/src/stat.cpp#L1792.
The problem arises in the following section:

 K = std::min(K, src2.rows);

    _dist.create(src1.rows, (K > 0 ? K : src2.rows), dtype);
    Mat dist = _dist.getMat(), nidx;
    if( _nidx.needed() )
    {
        _nidx.create(dist.size(), CV_32S);
        nidx = _nidx.getMat();
    }

K is reduced, when test.jpg (src2) is inspected, since it only has 9 rows.
_nidx.create(dist.size(), CV_32S); reduces the dimension of nidx to 9 columns, invalidating older results.
Later images may increase the column size again to 10.
The assumption, that the number of descriptors of every image is always > nn, seems unenforced.