OpenCV Mat object copying speed up - c++

Question

Recently i switched from opencv-python to c++ version of opencv, because i want to speed up my real time video-processing app with CUDA. I am new to C++ so i found some unclear moments with memory management while optimizing my code.
For example, i have some filter chain like this:
void apply_blue_edgess(cv::Mat& matrix, cv::Mat& mask, cv::Mat& inverted_mask) {
cv::Mat gray_image, blured, canny, canny_3d, in_range_mask;
cv::cvtColor( matrix, gray_image, CV_BGR2GRAY );
cv::GaussianBlur( gray_image, blured, cv::Size( 5, 5 ), 0, 0 );
cv::Canny(blured, canny, 0, 100);
cv::cvtColor( canny, canny_3d, CV_GRAY2BGR );
cv::inRange(canny_3d, cv::Scalar(255,255,255), cv::Scalar(255,255,255), in_range_mask);
canny_3d.setTo(cv::Scalar(0, 171, 255), in_range_mask);
cv::GaussianBlur( canny_3d, matrix, cv::Size( 5, 5 ), 0, 0 );
cv::bitwise_and(matrix, mask, matrix);
}
Is it ok to use new Mat object at every step of the filter chain (gray_image, blured, canny, canny_3d, in_range_mask) ? Is such continuous memory allocation bad for performance? If so, how should i write similar functions?
As was suggested in the comment section, i ended up doing functor wrapper:
struct blue_edges_filter {
blue_edges_filter(int width, int height)
: gray_image(width, height, CV_8UC1),
blured(width, height, CV_8UC1),
canny(width, height, CV_8UC1),
canny_3d(width, height, CV_8UC3),
in_range_mask(width, height, CV_8UC3)
{ }
int operator()(cv::Mat& matrix, cv::Mat& mask, cv::Mat& inverted_mask) {
cv::bitwise_and(matrix, mask, internal_mask_matrix);
cv::bitwise_and(matrix, inverted_mask, external_mask_matrix);
cv::cvtColor( matrix, gray_image, CV_BGR2GRAY );
cv::GaussianBlur( gray_image, blured, cv::Size( 5, 5 ), 0, 0 );
cv::Canny(blured, canny, 0, 100);
cv::cvtColor( canny, canny_3d, CV_GRAY2BGR );
cv::inRange(canny_3d, cv::Scalar(255,255,255), cv::Scalar(255,255,255), in_range_mask);
canny_3d.setTo(cv::Scalar(0, 171, 255), in_range_mask);
cv::GaussianBlur( canny_3d, matrix, cv::Size( 5, 5 ), 0, 0 );
cv::bitwise_and(matrix, mask, matrix);
}
private:
cv::Mat gray_image, blured, canny, canny_3d, in_range_mask;
};
//Usage
blue_edges_filter apply_blue_edgess(1024, 576);
apply_blue_edgess(matrix, mask, inverted_mask);

Answer 1

You can reuse memory without allocation. Create temporal images:
void apply_blue_edgess(cv::Mat& matrix, cv::Mat& mask, cv::Mat& inverted_mask)
{
cv::Mat tmp[2];
int srcInd = 1;
auto InvInd = [&]() -> int { return srcInd ? 0 : 1; };
cv::cvtColor( matrix, tmp[InvInd()], CV_BGR2GRAY );
srcInd = InvInd();
cv::GaussianBlur( tmp[srcInd], tmp[InvInd()], cv::Size( 5, 5 ), 0, 0 );
srcInd = InvInd();
cv::Canny(tmp[srcInd], tmp[InvInd()], 0, 100);
srcInd = InvInd();
cv::cvtColor( tmp[srcInd], tmp[InvInd()], CV_GRAY2BGR );
srcInd = InvInd();
cv::inRange(tmp[srcInd], cv::Scalar(255,255,255), cv::Scalar(255,255,255), tmp[InvInd()]);
tmp[srcInd].setTo(cv::Scalar(0, 171, 255), tmp[InvInd()]);
cv::GaussianBlur( tmp[srcInd], matrix, cv::Size( 5, 5 ), 0, 0 );
cv::bitwise_and(matrix, mask, matrix);
}