Question

Given two cv::Mat matrices that maps every pixel in the source image to a pixel in the destination image (R2 to R2), I would like to transform a source image to a destination image. I’ve sucessfuly done so using for loops but it is too slow:

cv::Mat srcImg(100,100,CV_8U);

//fill...

cv::Mat dstImg(100,100,CV_8U);

//dst2src ->backprojection

//these matrices indicates for each pixel in the destination image, where to map it from the source image

cv::Mat x_dst2src(100,100,CV_64F);

cv::Mat y_dst2src(100,100,CV_64F);

//fill...

for(int ydst=0; ydst!=100;++ydst)

{

    for(int xdst=0; xdst!=100;++xdst)
     {
      double xsrc = x_dst2src.at<double>(ydst,xdst);
      double ysrc = y_dst2src.at<double>(ydst,xdst);
      double val = getBicubic(srcImg,xsrc,ysrc);
      dstImg.at<double>(ydst,xdst) = val;

     }

}

this basic code works, but VERY slow (my images are bigger than 100×100, and I must use bicubic).
Thanks,

-O-

Answer 1

There’s a nice OpenCV function that’s called remap(). And it does, surprisingly, exactly this transform.

To speed it up, you should look for another function that prepares the maps in a format that’s a bit more processor-friendly than a simple position map. (It’s something like mapTransform(), look for it in the see also part of the remap() docs)

Happy remapping!