On 5/5/2019 1:16 PM, Even Rouault wrote:
Glenn,
I'm trying to track down a discrepancy between GDAL 1.7.2 and GDAL
2.4.0. The NASA Worldwind code makes a call to gdal.ReprojectImage()
using a dataset generated from an input file to a working destination
dataset that's empty, except that the alpha band has been filled with
0xFF. In the case I'm examining, the projection is 1 to 1; no change.
The source file has an appropriate alpha mask to clip the image so only
relevant data is shown.
When using the 1.7.2 release, the image is shown with a proper overlay
on the lower priority layers, with the portions with alpha mask of 0
being completely transparent. But the more recent 2.4.0 release has a
black boundary around the image, indicating that the alpha mask was set
to 0xff for the whole image.
It looks like there's a difference in behavior in the Java call
gdal.ReprojectImage(srcDS, destDS, s_srs_wkt,
t_srs_wkt, gdalconst.GRA_Average);
For the 1.7.2 release, the alpha mask in the destination data set is
filled in correctly, matching the alpha band in the source. But in the
2.4.0 release, the destination alpha band is filled with 0xFF, which
sort of explains the black boundary. Curiously, for the 1.7.2 release,
it doesn't matter if the destination alpha band was filled with zeros or
0xFF. For the 2.4.0 release, if the destination dataset alpha band is
filled with zeros, the image displays correctly, and the reprojected
destination dataset alpha band matches the source.
I'm trying to understand what the purpose is of the destination alpha
density mask in 'gdalwarpper.cpp'. In the file 'gdalwarpoperation.cpp',
I see a note that says "Note that detailed semantics of the masks should
be found in
GDALWarpKernel." And in 'gdalwarpkernel.cpp': "A detailed explanation
of the semantics of the validity and density masks, and their effects on
resampling kernels is needed here."
So if there's a place where the semantics are explained, please send me
link. And, if this rings a bell about a change that was made some time
between the 1.7.2 (April 23, 2010) and the 2.4.0, please let me know.
Looking at the NEWS file, the change of behaviour likely dates back to GDAL
1.8.0:
* GDALReprojectImage() : correctly assign nSrcAlphaBand and nDstAlphaBand
(#3821) --> https://trac.osgeo.org/gdal/ticket/3821
Previously the source and target alpha bands were treated as regular bands.
If the target alpha band is initialized to 0xFF, then it means that all target
pixels are valid, and thus they will remain the same if there is no
corresponding source pixel, or if they have source_alpha = 0.
The maths involved with destination density are in:
https://github.com/OSGeo/gdal/blob/master/gdal/alg/gdalwarpkernel.cpp#L1330
https://github.com/OSGeo/gdal/blob/master/gdal/alg/gdalwarpkernel.cpp#L1414
Basically this is for alpha blending of source onto target.
That said, looking at the above mentionned change in GDALReprojectImage(), I
noticed there was something slightly wrong, which I just fixed, in which the
alpha bands were treated both as regular bands and alpha bands. So I'm not
sure what the effect was exactly. That said, the difference in results seemed
minimal when warping over a target image initialized with alpha = 0.
So all in all, you should initialize your target image with alpha = 0 as you
noticed.
Even
It turns out that the change you made for treating the alpha band
separately
(https://github.com/OSGeo/gdal/commit/b427cb23ff1bc2a5ba2f6634fc3bd90816bb7789)
fixed one other problem that you might have intended to fix, but has
caused another.
I have a general question - the new code assumes that the alpha band
will always be the last one in the dataset. Is that guaranteed by some
standard? In the previous code, the alpha band could have been band 1,
and the band number was recorded for later use.
After I changed the WorldWind code to initialize the target alpha band
to 0, which fixed my immediate problem, I discovered that if the source
dataset has no alpha band, that the alpha band of the result has all
zeros. That doesn't happen with your change - it doesn't matter what's
in the target alpha band initially if there is no alpha band in the
source - after reprojectImage, the alpha band of the result has all 255
regardless of the values initially in the target dataset. That makes sense.
Perhaps that's the problem you intended to fix.
The new problem in the WorldWind code is that 'reprojectImage' was used
to clip the source dataset to a 'region-of-interest' (ROI). That's
probably an improper use of reprojectImage. If you refer to this code:
https://github.com/WorldWindEarth/WorldWindJava/blob/develop/src/gov/nasa/worldwind/data/GDALDataRaster.java
in createMaskDataset at line 466, it's creating a dataset with a single
alpha band for the ROI. When the mask dataset is used at line 1135 in a
reprojectImage call, GDAL correctly reports a warning:
GDAL error 5 : GDALWarpOptions.Validate(): nBandCount=0, no bands
configured!
The intent was to generate a dataset with alpha=0 outside of the ROI.
In the call to GDALUtils.composeDataRaster at line 1154, the alpha band
is retrieved from the mask data set, and eventually used in
'GDALUtils.applyImageMask()' to clip the image.
https://github.com/WorldWindEarth/WorldWindJava/blob/develop/src/gov/nasa/worldwind/util/gdal/GDALUtils.java
So in order to work with any GDAL release after 3.0.0, that code will
need to be reworked.
If you tell me that WorldWind's use of 'reprojectImage' is inappropriate
for this purpose, that's fine. It will need to be re-written. But
other users might be using 'reprojectImage' for similar purposes, other
code might break. Some entry in the release notes might help.
_______________________________________________
gdal-dev mailing list
gdal-dev@lists.osgeo.org
https://lists.osgeo.org/mailman/listinfo/gdal-dev
