Hi,
we encountered a problem with box filtering of float images.
If the image only has relevant image content in the middle and is otherwise 0, there are errors on the right and bottom, which are probably caused by the update process.
The filter response should also be 0 in these areas.
The problem causes subsequent errors in our applications with natural images.
The problem can be reproduced with the following sample code:
void main()
{
// Boxfilter auf float-Bildern
IppiSize maskSize = { 55, 55 };
IppiSize imageSize = { 1000, 1000 };
Ipp32f* psrc=nullptr;
Ipp32f* pDst = nullptr;
Ipp8u* pBuffer = nullptr;
int len = imageSize.height * imageSize.width;
// Create and fill source image
psrc=ippsMalloc_32f(len);
ippsZero_32f(pSrc, len);
IppiRect fillArea = { 400, 400, 200, 200 };
Ipp32f fillValue(1000.67);
for (int y = fillArea.y; y <= fillArea.y + fillArea.height; y++)
{
ippsSet_32f(fillValue, pSrc + y * imageSize.height + fillArea.x, fillArea.width);
}
// Create dest image
pDst = ippsMalloc_32f(len);
// Get work buffer size
int bufSize(0);
if (ippiFilterBoxBorderGetBufferSize(imageSize, maskSize, ipp32f, 1, &bufSize))
throw("Init failed");
pBuffer = ippsMalloc_8u(bufSize);
// Filter the image
Ipp32f borderValue(0);
int lineStep = sizeof(Ipp32f) * imageSize.width;
if (ippiFilterBoxBorder_32f_C1R(pSrc, lineStep, pDst, lineStep, imageSize, maskSize, ippBorderRepl, &borderValue, pBuffer))
throw("Filter failed");
if (pBuffer) ippsFree(pBuffer);
if (pSrc) ippsFree(pSrc);
if (pDst) ippsFree(pDst);
}
Result for l9 scaled with maximum 1e-5 (Name: ippIP AVX2 (l9), version: 2021.7 (r0xa954907f), build date: Nov 4 2022, cpu: l9):
Result for k0 scaled with maximum 1e-5 (Name: ippIP AVX-512F/CD/BW/DQ/VL (k0), version: 2021.7 (r0xa954907f), build date: Nov 4 2022, cpu: k0):
We have same results with versions 2020.0.1 and 2021.7.
Kind regards, Roy
Link kopiert
Hi Shanmukh,
I have created the image views again with marking of the relevant areas.
The source image pSrc is created and filled in lines 11 to 21. This image looks like this:
After filtering I get the following image (This looks ok without scaling
With scaling, however, you can see stripes outside of the expected filter response (in this picture for k0):
At l9, the stripes below and to the right of the valid filter response are even larger (see first image of the original query).
Kind regards,
Roy
