c# - Drawing match pairs in image-calculated by KNN and the potential bug in Features2DToolbox.DrawMatches -

i wrote code finds k closest matches knn algorithm. after getting matmatch , matchindices tried draw match pairs between 2 consequence frames.

i feed matmask , matchindices function features2dtoolbox.drawmatches :

image<bgr, byte> imresult = features2dtoolbox.drawmatches(immodelcurr, immodel.keypoints, imobserprev,imobser.keypoints, **matchindices**, new bgr(system.drawing.color.yellow), new bgr(system.drawing.color.red), **matmask**, features2dtoolbox.keypointdrawtype.not_draw_single_points); 

http://www.emgu.com/wiki/files/2.4.0/document/html/e92d37e6-fe4a-ad09-9304-cd2d2533bfa8.htm noticed gives me wrong drawing between matching pairs:

then tried implement such function myself:

 (int = 0; < matmask.rows; ++i)         {             if (**matmask[i, 0]** > 0)              {                 int indforcurrfrm = **matchindices[i, 0]**;                 int indforprevfrm = i;                  //for frame i-1                 pointf fromfirstframe = getimgobserved(keypoints[indforprevfrm]);                  //for frame                 pointf nextcorrespondingmatchedframe = getimmodelxy(keypoints[indforcurrfrm]);                  imcolorprv2.draw(new circlef(fromfirstframe, 5), new bgr(mtchcolor), 3);// frame i-1                 imcolorshow.draw(new circlef(nextcorrespondingmatchedframe, 5), new bgr(mtchcolor), 3); // frame                   // draw line on own matching               imresult.draw(new linesegment2df(fromfirstframe,nextcorrespondingmatchedframe),new bgr(system.drawing.color.floralwhite),1);              }         } 

and fetch corresponding pair point coordinates (x,y) , draw myself [ see results in snapshot].

one left bottom see matches(shown white line) , each corresponding pairs circle in same color [by own function] , on other side-bottom right , results drawn drawmatches function emgu.pay attention these 2 functions use same matmash , matchindices.

so wondering if drawmatches @ emgu has bugs or doing somewhere wrong?

it might useful filter best matches matchindices, using euclidean distance between descriptors of each corresponding pair . times 1 keypoints first image can match many keypoints second image, can confuse plotted lines in results.

filtering like:

vector< dmatch > filtered_matches;   for( int = 0; < descriptors_of_model.rows; i++ )     {        if( matchindices[i].distance < small_threshold )           filtered_matches.push_back( matchindices[i]);        }