Re:We Built a Facial Recognition Machine for $60
by bsenftner on Tuesday April 16, 2019 @11:16AM
(#19674357)
Bingo. 1) Train an algorithm that given any angle view of a human head, produce a 3D reconstruction of said head. (Depending upon the amount of face visible, the accuracy of the reconstruction increases. 2) Train an algorithm that given a human head geometry, it identifies any deformation due to facial expression and neutralizes that expression. 3) Trivial algorithm: given a human head geometry not facing the camera, rotate that head such that it is facing the camera. 4) render the human head geometry without expression, facing the camera, using the video image source as the texture map for said head. 5) Perform FR.
I think you've missed the last 5-ish years of NN development. Some of the latest Face2Vec stuff can handle different perspectives pretty well by either building a system that can predict 3D representations of the face from one shot[0] or by just training on two different perspectives of the same face and having the system output the same vector (I've read some work on this but Google is failing me at finding a reference).
If you can recognize people from different angles, computers will eventually be able to do it better than you. Probably already.
[0]https://arxiv.org/pdf/1703.10714.pdf