@MaximSmolskiy, thank you for the patch. It looks like those matrix expression-based calculations are often done on single-element scalars, and the only reason why cv::Scalar and div are used is because cv::sum() returns a scalar, rather than, say, double. I wonder, if we change mask_sum definition to:

double mask_sum = sum(mask)[0];

maybe the patch can be further simplified and some of matrix expressions could be replaced with simple arithmetic operations (which is much faster)?

thank you for the patch. It looks like those matrix expression-based calculations are often done on single-element scalars, and the only reason why cv::Scalar and div are used is because cv::sum() returns a scalar, rather than, say, double. I wonder, if we change mask_sum definition to:

double mask_sum = sum(mask)[0];

maybe the patch can be further simplified and some of matrix expressions could be replaced with simple arithmetic operations (which is much faster)?

@vpisarev Then this will probably require adding additional branches, because in some cases mask_sum has several different elements:

[ RUN      ] MultiChannelMask/Imgproc_MatchTemplateWithMask.CompareNaiveImplCCOEFF/0, where GetParam() = (32FC3, 32FC3)
mask_sum = [678.275, 646.07, 686.876, 0]