gdaladdo: cubic resampling produces bad mask

[cdestigter]

Given an RGB+mask image, running gdaladdo with cubic resampling computes masks in a way that some originally masked pixels aren't masked in overviews. This results in random black pixels around the edges of data areas.

The output from average and nearest seems correct, so it seems cubic is doing something strange here.

Steps to reproduce:

for r in cubic nearest average
do
    echo $r 
    cp masked.tif $r.tif
    gdaladdo --config INTERLEAVE_OVERVIEW PIXEL --config GDAL_TIFF_INTERNAL_MASK YES --config GDAL_CACHEMAX 128 -r $r $r.tif 2 
    tiffsplit $r.tif $r-
done

The attached zip includes a qgis project which highlights the differences between the algorithms.

[Even Rouault]

Grrr, my comment got lost.

Craig, I can see the issue in cubic-aac.tif, but when I try to reproduce, the result I get looks fine. Are you using latest GDAL trunk ? Which compiler, environment, compilation flags are you using ?

[cdestigter]

You're right, I'm not seeing the bug now either. I think I was running this on the wrong machine and it had an older GDAL than I thought it did. Sorry for the noise.