I write facial recognition software. If those example facial images and their match scores are any indication of their systems accuracy, it is on the poor side.
The point is that they made it available to customers as a full, scalable pipeline. They can (and will) improve their accuracy continuously. Any kiddo can take some bleeding-edge wide resnet variation and get better results than they have right now, but can't make it available at scale to anyone.
"but can't make it available at scale to anyone." =- I guess that's the popular misconception. Their system scales awful, rendering it's expense of operation on the high end side, for a weak product. They are simply exploiting their name.
