(Not really a) Problem with nppiFilterRow function with kernel that has negative values

Accelerated Computing GPU-Accelerated Libraries
1

I am using the nppiFilterRow_32f_C1R( )function with a kernel defined as:

{0.0f, -0.0229f, -0.0418f, -0.0477f, -0.0344f, 0.0f, 0.0516f, 0.1114f, 0.1670f, 0.2065f, 0.2207f),
0.2065f, 0.1670f, 0.1114f, 0.0516f, 0.0f, -0.0344f, -0.0477f, -0.0418f, -0.0229f, 0.0f}

The size of the kernel is 21 and the radius is 10. The call looks like:

nRet = nppiFilterRow_32f_C1R(md_curFrameP, (448sizeof(float)), md_buf1P, (448sizeof(float)), {448, 448}, md_kernelP, 21, 10);

where
md_kernelP points to the array shown above that has been copied to device memory.
md_curFrameP points to a 448x448 float image in device memory where the values are between 0.0f and 1.0f.
md_buf1P points to 448x448 float image where the result will be written.

The input image is essentially a black frame with a single vertical non-black stripe in the center. When I run the above filter the result is a filtered with white bands on either side. If I replace the negative components with 0.0f, I get just the filtered stripe I expected.

Original image: [url]https://imgur.com/2aLb4eP[/url]
Filtered image with negative kernel values: [url]https://imgur.com/8uDjYLZ[/url]
Filtered image without negative kernel values replaced with 0.0f: [url]https://imgur.com/1rECxL0[/url]

The kernel values used are derived from the following matlab code:

h = sinc([-MATCH_SIZE:MATCH_SIZE] / MATCH_SIZE * 2);                     % Create Filter Kernel
h = h / sum(h);

where MATCH_SIZE is 10 and h is a 21 element vector.

Why do the negative kernel values cause the white bands?

2

because the negative values in the kernel mean you can have negative values in your output image.

The question then becomes what pixel characteristics to assign to a negative valued pixel. Apparently it is bright white.

3

OMG…it seems perfectly obvious when someone “says it out loud”. And, of course, I was clipping negative values to zero in the kernel that I was comparing to.

Thank you.