To figure out how this translates to recognition -- "who is this person?" -- you can roughly take the verification rate V and number of different people in your database N and get a recognition rate of roughly V^sqrt(N). (This is a very very rough estimate!)
So with 1 person in your database, your rate is just V = ~90%. With 100 people, it's already down to 35%. And so on...
(I'm the author of one of the best methods on LFW right now -- the attribute-based one [2].)
[1] http://vis-www.cs.umass.edu/lfw/
[2] http://www.cs.columbia.edu/CAVE/projects/faceverification/
I think the vision community is thus looking for how to exploit context effectively. There have been a number of interesting approaches recently, from looking at "things" vs "stuff", to using additional information such as transcripts and captions, etc.
And of course, there are still lots of people looking at how to better solve the basic recognition problem itself.
Here is an idea to validate my concern:
1. Crawl over all photos on Facebook and run a frontal-face detector with it;
2. Use current state-of-art methods (attribute based method or hybrid methods) to get several pairs of similar faces (highest score ones) of different people;
3. Use the same method as in step 2 to get several pairs of similar faces of same person with roughly the same score as in the step 2;
4. Have human volunteers to blindly judge which pairs of person are in fact the same one given only a small region of picture (e.g. face should occupy 20% of the total area);
5. Check the result in step 4 with ground truth data;