Why is the quality adjustment restricted to low similraity scores?

[SebastianEngelhardt]

This is more a question for the paper instead of the code (I hope this is still fine). You write:

A high similarity score s can only be achieved through the comparison of two high-quality samples. In this case, the similarity is well reflected in the comparison score s and thus, no quality-based score adjustments (ω = 0) is needed.

I would actually doubt that initial assumption. We see high similarity scores for two images with both have low quality. See for example here https://pages.nist.gov/frvt/html/slides/multiple_faces_single_image.pdf the "Sam vs Yao" comparison.

Did you do some experiments with a function ω that also adjust those cases?