Hello all,
I think I may have found a bug with the ImageBufAlgo::nonzero_region under
some specific circumstances, but I wanted to lay it out here
before opening a Github issue, just in case this is (or would be)
considered a known bug or “wontfix” edge-case.
To start with, let's manufacture an EXR with mixed channel formats:
oiiotool --pattern black 640x480 3 --chnames "R,G,B" --pattern fill:left=0.1,0.1,0.1:right=0,0.75,0 640x480 3 --chnames "foo.R,foo.G,foo.B" --chappend -d half -d foo.R=float -d foo.G=float -d foo.B=float -o multi_dtype.exr
From there, I use the Python bindings to read the image and compute the non-zero ROI:
buf = OpenImageIO.ImageBuf('multi_dtype.exr')
nonzeroROI = OpenImageIO.ImageBufAlgo.nonzero_region(buf)
print(nonzeroROI)
This prints the expected result: 0 640 0 480 0 1 0 6
However, if I do the same thing but force a read of the image before computing the non-zero region, I get an incorrect result:
buf = OpenImageIO.ImageBuf('multi_dtype.exr')
buf.read(force=True)
nonzeroROI = OpenImageIO.ImageBufAlgo.nonzero_region(buf)
print(nonzeroROI)
In this case, the printed result is: 0 640 0 360 0 1 0 6
Now, I understand that ImageBuf.read is
most likely promoting the half channels to float so they can all
be stored in a single buffer, which is fine, but the fact that the
subsequent nonzero_region returns
an incorrect result definitely seems like a bug (e.g. the IBA
function doesn’t “know” that the underlying data for the half
channels has been promoted to float).
Other notes:
- This is using OIIO 2.2.19.0. I
looked at the 2.3.11.0 source for ImageBuf, and it didn’t seem
like the forced-reading code had changed significantly, but
there’s a chance this is hitting a corner-case that has been
addressed already in newer versions.
- I don’t have a build of 2.3 handy to test, so I need to try to get one built, but I’m wondering if maybe you can confirm whether something might have changed that could affect this in the meantime.
- I’m not sure whether this issue is specific to the file structure and pixel data in my example (half-float black RGB channels, with additional non-zero float channels).
- I’m not sure if this is just a bug
with
ImageBufAlgo::nonzero_region, or if this could point to a more widespread source of unexpected behaviors when usingImageBufAlgoon ImageBufs with “local” storage.
Let me know if this is in fact a bug and I'll open a proper issue for it.
Thanks for any insights.
-Nathan
_______________________________________________ Oiio-dev mailing list Oiio-dev@lists.openimageio.org http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org
