I don't know about face recognition but I've quantatively analysed their speech recognition and it came dead last after the 5 or so others that I tested.
I know others have asked, but I'd also be interested in that ASR analysis. I'm working on something for my masters involving crowd sourced transcribers and we have access to Watson's ASR, which the (non CS) administration likes, but I've suspected its crap and would be cool to have some data.
Yeah, this would be good to share with the team if you have it documented.