How does this implementation perform compared to classic approaches such as OpenALPR [1] ?

At the very least running local inference becomes much more expensive, and possibly provides worse results.

[1] - https://github.com/openalpr/openalpr