recently updated? currently a terrible matching algorithm
it looks like there was an update recently where with my own testing the accuracy of matching of a known baseline faces went from >=99%, as expected since it it is in fact the same person using two similar facial poses in the same photo set, to now in the mid 70s to 80s percentiles. I've tested across multiple known baselines that had a high degree of likeness and similar results appear.
This could be that new algorithm is much more discerning when assigning a high score match, which would be okay, however you'd also expect a wide % difference between two images of different people, but that isn't the case. Comparing several datasets against a few baseline tests shows a relatively narrow % difference between same person and similar looking but different people.
This was an extremely accurate and useful tool but I'm having a hard time making sense of the new scoring spectrum from different to similar to actual same person matches.