The ippiFilterGaussianBorder puzzles me, notably the sigma parameter. Compared to doing the Gaussian Blur with a matrix, ippiFilterGaussianBorder gives me what looks like unprecise results. This is with a typical kernel size 15 or 31. Of course, a Gaussian Blur is separable, so we can replace the matrix with two one-dimensional matrices.
Can you give an example of a matrix or two separable matrices where ippiFilterGaussianBorder is (or should be) equivalent to ippiFilterBorder ?
for kernelSize = 15 and sigma = .5.5 coefficients are the next:
k0 = 0.038978182
k1 = 0.048321572
k2 = 0.057956705
k3 = 0.067252662
k4 = 0.075501995
k5 = 0.082006926
k6 = 0.086175894
k7 = 0.087612126
k8 = 0.086175894
k9 = 0.082006926
k10 = 0.075501995
k11 = 0.067252662
k12 = 0.057956705
k13 = 0.048321572
k14 = 0.038978182
the filter itself is separable - therefore row/column coefficients are the same. The only 1 fact that is not reflected in the documentation - that these coefficients are normalized on their initial sum (coeff are obtained according to the formula in the manual and then divided by their sum).
