BFMatcher with Hamming distace stores distances as integers instead of floats (Patch #2245)

When BFMatcher is used with Hamming distance, BFMatcher::radiusMatchImpl() first calculates the result as integer to Mat dist. Then they are converted into float values and stored in Mat distf.

However, at http://code.opencv.org/projects/opencv/repository/revisions/master/entry/modules/features2d/src/matchers.cpp#L429 original dist matrix is still used to fetch the values instead of correct distf matrix. Integer values are just reiterpreted to be floats: const float* distptr = dist.ptr<float>(qIdx);

For loop at line http://code.opencv.org/projects/opencv/repository/revisions/master/entry/modules/features2d/src/matchers.cpp#L432 should probably be updated to use distf for consistency as well as it uses original dist matrix: for( int k = 0; k < dist.cols; k++ )

Current bug breaks radius searches and visualizations of matches point sets when Hamming distance is used.