Colored patches regression in Verdandi

[Kevin Mehall]

I'm occasionally seeing bad results from Verdandi in which parts of the panorama are discolored:

This looks similar to an issue previously reported and fixed in 2019: https://groups.google.com/g/hugin-ptx/c/mpnWNfJZ4b8/m/LMFVNMdbBQAJ

In testing, I found that 2023.0.0 and mercurial tip are affected, while 2022.0.0 results are good.

Weirdly, it seems that the nona output is nondeterministic for the same images and PTO file. The nona output itself looks fine, and the artifacts are introduced by verdandi. The severity and coloring of these artifacts varies between different sets of remapped images from running nona multiple times, but does not seem to change between different runs of verdandi on the same remapped TIFs.

There are only a few commits touching verdandi between 2022.0.0 and 2023.0.0, but 8556:17d25b5ae116 stands out. It re-adds a call to Multigrid that had been commented out with the note "this results in artefacts, maybe there is still a small error in other lines" in a 2019 change that I think was the bugfix from the thread linked above.  I tested commenting out that line again and the results look better.

I don't understand the algorithm enough to know how this helps, what other problem it might be working around, or what the proper fix would be, but I'm hoping someone can take another look at it.

Here are two sets of remapped images that exhibit the problem:
https://test-bucket.stscn.net/pano/verdandi-color-bug/1_remapped.zip
https://test-bucket.stscn.net/pano/verdandi-color-bug/2_remapped.zip

Run verdandi --wrap --seam=blend --compression=LZW -o out.tif -- part*.tif.

Let me know if I can help with more details, testing, or anything else.

Thanks,

Kevin

[Kevin Mehall]

Image was removed from the post, trying again: