image-similarity-measures VS COMET

Speaking of math, using this library -- https://github.com/up42/image-similarity-measures -- I computed the following for these images vs the original image:

The easiest place to start is using the classical approaches such as implemented here. For the kind of qualitative assessments you're performing, you'd probably need to use some deep learning techniques but these generally require significant technical background to implement.

I use https://github.com/up42/image-similarity-measures for image similarity.

It's definitely a hard task to evaluate. I think we can use models like https://github.com/Unbabel/COMET for translation to try and mimic human evaluation. I don't know if datasets exist for that. There are some research done about that : https://aclanthology.org/P19-1502/ https://arxiv.org/abs/2104.00054v1

2/ Evaluation We compare Open AI to DeepL, ModernMT, NeuralSpace, Amazon and Google. A lot of metrics exist for automatic machine translation evaluation. We chose COMET by Unbabel (wmt21-comet-da) which is based on a machine learning model trained to get state-of-the-art levels of correlation with human judgements. (read more on their paper ) .

I'm trying to understand how I can use COMET to evaluate translation models https://github.com/Unbabel/COMET ? I don't really understand how it was trained the meaning of the outputed values ? https://unbabel.github.io/COMET/html/faqs.html#which-comet-model-should-i-use