The face engine work best when both eyes of the person are visible.
The single person view (Find Similar Faces or double click on a confirmed/unnamed faces) is a more efficient way to find and label similar looking unnamed faces at once. It is good method to label a few of the known faces that you are interested in at once.
The face detection is applied to specific zones of a photo. If you have for example a face in side-profile, it's possible to drag your own box to define a name. This video gives more information on the new feature:
My experience so far with LRs face detection is a little disappointing as well.
I understand that machine face recognition has severe limits. Persons age, change in expression, glasses, hats... not to speak of angle of shooting, light conditions etc etc. Plus it needs to be computationally feasible.
And yet when I started to name a person, particularly after having tagged some childhood pictures, the suggestions started to get very very odd, including faces from the opposite sex, some graphic ornaments on a wall, and ancient Greek statues.
And then going through a large collection (again!), with thousands of clicks...
Adding to the disappointment: Although LR CC writes face data to XMP-mwg-rs tags (jay!), it does NOT read and process existing XMP-mwg-rs written by other software (Nooooooo!).
Thousands and thousands clicks to go, again...
I understand the inherent limitations, but not why a photo like the ones below aren't having faces detected when they are so obviously faces. I've had it detect half faces (only forehead and eyes visible) in random photos, which makes this especially puzzling. And like I say, no obvious way of making it recheck photos even after spending some time tagging.