Terrible, unless you're in a "cooperative" setting, i.e., where the user is in controlled lighting conditions and looks directly at the camera, in a decent-to-high res photo.
Labeled Faces in the Wild (LFW) is the de-facto standard for looking at real-world face verification ("are these two faces of the same person?"): http://vis-www.cs.umass.edu/lfw/results.html
The top performer is currently around 93% (random chance is 50%). For a very rough guide on how recognition rates would be ("who is this person?"), it's v^sqrt(N), where v is the verification rate and N is the number of different people you're trying to distinguish between. Note that it goes down very fast with N.
Labeled Faces in the Wild (LFW) is the de-facto standard for looking at real-world face verification ("are these two faces of the same person?"): http://vis-www.cs.umass.edu/lfw/results.html
The top performer is currently around 93% (random chance is 50%). For a very rough guide on how recognition rates would be ("who is this person?"), it's v^sqrt(N), where v is the verification rate and N is the number of different people you're trying to distinguish between. Note that it goes down very fast with N.