c# - OpenCV MatchTemplate on GPU -

in c# project need fast template matching algorithm. have opencv implementation. simplified code is:

 using opencvsharp;  using opencvsharp.cplusplus;  // ...   var image = new mat("image.png");  var template = new mat("template.png");   double minval, maxval;  point minloc, maxloc;  var result = image.matchtemplate(template, matchtemplatemethod.ccoeffnormed);  result.minmaxloc(out minval, out maxval, out minloc, out maxloc);  console.writeline("maxloc: {0}, maxval: {1}", maxloc, maxval); 

i'd same calculations on gpu speed them significantly, cannot find example of how that. example on github commented out , doesn't seem working.

i've attempted port opencv's matchtemplate() function gpu before.

they give equations in their documentation method. called "normalized cross-correlation."

the naive approach port gpu (i did pixel shader). turned out slower opencv's cpu method. why? using method called "fast normalized cross-correlation" outlined in this paper j.p lewis ilm.

the trick correlation in frequency domain , take advantage of convolution theorem: point-wise multiplication in frequency domain equivalent convolution in spatial domain. means time complexity reduced, , have add fft.