GDALResampleChunk32R_Gauss should use optimal antialiasing kernel

[wkarel]

GDALResampleChunk32R_Gauss uses the following 3x3 kernel for scale factors of up to 2:

1,2,1,
2,4,2,
1,2,1

While this kernel would be optimal (for the purpose of antialiasing) for a scale factor of sqrt(2), it is not for the common scale factor of 2 commonly used to build overviews (hierarchically). For a scale factor of 2, the 5x5 kernel defined in GDALResampleChunk32R_Gauss would be optimal, but is not used in that case, unfortunately. As there is no way to force usage of a certain kernel, I consider this a bug - specifically, because in ticket #2137, the main purpose of implementing the Gaussian filter was antialiasing.

Interestingly, it seems that the source data read by GDALRegenerateOverviewsMultiBand seems to already consider this, as it multiplies nKernelRadius (which is 1 for GAUSS) with nOvrFactor. Hence, only the choice of kernel in GDALResampleChunk32R_Gauss would need to be changed:

if(nResYFactor <= 2 )
{
  panGaussMatrix = anGaussMatrix5x5;
  nGaussMatrixDim=5;
}

For the optimal standard deviation of a Gaussian kernel for downsampling by a factor of 2, see e.g. Lowe, 2004: "Distinctive Image Features from Scale-Invariant Keypoints" or the OpenCV-documentation on pyrDown: ​http://docs.opencv.org/3.1.0/d4/d86/group__imgproc__filter.html#gaf9bba239dfca11654cb7f50f889fc2ff

[Even Rouault]

This ticket has been automatically closed because Trac is no longer used for GDAL bug tracking, since the project has migrated to GitHub. If you believe this ticket is still valid, you may file it to ​https://github.com/OSGeo/gdal/issues if it is not already reported there.