image - Difference of cv::filter2D and self-contained convolution on Gabor filter -

due reason, have write filter function myself. following convolution function.

void convolve (cv::mat& f, cv::mat& w, cv::mat& output) {   output = f.clone();    int height = f.rows;   int width = f.cols;   int = (w.rows - 1) / 2;   int b = (w.cols- 1) / 2;    cv::mat f2 = f.clone();    (int x = a; x < height - a; ++x)   {     (int y = b; y < width - b; ++y)     {       float sum = 0.0;       (int s = -a; s <= a; ++s)       {          (int t = -b; t <= b; ++t)         {           sum += w.at<float>(s+a, t+b) * f2.at<float>(x+s, y+t);         }       }       output.at<float>(x, y) = sum;     }   } } 

then compare using function , cv::filter2d function, , discover different after filtering. filtered image same when theta , psi 0 not others.

int main() {     cv::mat in = cv::imread("something.jpg", 0);     cv::mat dest, dest1;     cv::mat src_f;     in.convertto(src_f, cv_32f);      int kernel_size = 31;     double sig = 1.0, th = 0.2, lm = 1.0, gm = 0.02, ps = 0.0;      cv::mat kernel =    cv::getgaborkernel(cv::size(kernel_size,kernel_size), sig, th, lm, gm, ps);      cv::filter2d(src_f, dest1, cv_32f, kernel);     convolve(src_f, kernel, dest);      cv::mat viz;     dest.convertto(viz, cv_8u, 1.0/255.0);     cv::mat viz1;     dest1.convertto(viz1, cv_8u, 1.0/255.0);      imshow("my dest", viz);     imshow("k dest", viz1);     cv::waitkey(); } 

any suggestion self-contained convolution ? 0 padding part ignored.

you access kernel in convolution implementations matrix of floats, composed of doubles (ksize set cv_64f default, according doc). looks opencv smart enough detect , perform conversion/correct memory access during convolution of matrices different element types, that's why produces correct results.

in order fix snippet added line

kernel.convertto(kernel, cv_32f); 

this produces visually equivalent result (up padding).

another option explicitly set ktype cv_32f.

also, when working big kernels (~11x11 , larger) opencv's filter2d implementation performs convolution using dft-based algorithm (in frequency domain), may different results , timings.