Thank you for this interesting project. Creating a matching decision in a higher dimension is an innovative idea.
I know that it has been a few years, but maybe this is still relevant: you should consider submitting your work to the FATE benchmark by NIST, the gold standard for face recognition.
The academic benchmarks like AgeDB or LFW are very small and easily saturated. It would be very interesting to see the performance on larger "real world" data.