Problem is not all photos (profile photos especially) are close-ups of a person's face. Although they could use facial tagging and maybe come up with a composite? I'm probably wrong though.
You don't need all photos, you can run a basic face recognition algorithm like the ones used in cameras to identify good candidates and just filter all other photos.
Profile photos are good enough, and with the amount of selfies people are taking there should be more than enough candidates for a full facial recognition matching.
Also facebook isn't the only social network, LinkedIn profile pictures are usually much better for facial recognition Google+ profile pictures are also usually quite good because they crop your face into that silly circle.
We usually "region propose" and crop to a certain area (in this case the face area, usually at 256x256) then transform to align eye areas before passing to training. This is to standardize the data beforehand. I'm not sure if this lib does region proposal but you can easily write a pre-processor with openCV face plugins to identify face regions (if any, maybe your training image is a landscape not a face!) for cropping.
Yes, the processing pipeline first does face detection and a simple transformation to normalize all faces to 96x96 RGB pixels. Then each face is passed into the neural network to get a 128 dimensional representation on the unit hypersphere.
For a landscape, face detection would probably not find any faces and the neural network wouldn't be called.
And an image with multiple people will have many outputs: the bounding boxes of faces and associated representations.
Profile photos are good enough, and with the amount of selfies people are taking there should be more than enough candidates for a full facial recognition matching.
Also facebook isn't the only social network, LinkedIn profile pictures are usually much better for facial recognition Google+ profile pictures are also usually quite good because they crop your face into that silly circle.