FG-Net in megaface challenge 1

Thanks for your great work in face recognition. But I have something confusing to ask for help. In the devkit of the megaface evaluation challenge, we see the script _run_experiments.py_ adopts Joint Bayesian(JB) method to compute the probability for face verification. However, we found in this paper[MegaFaceCVPR16](https://megaface.cs.washington.edu/KemelmacherMegaFaceCVPR16.pdf) that the metric for comparison is L2 distance. So, we wonder which one we need to use, L2 directly or Joint Bayesian?