Hmm...
This is old research, which seems to be a recreation of the work of Sutcu et al. among others.
I did my masters thesis on this.
I feel like the ability for this method to work well depends on the methodology of taking the enrollment and the subsequent key-generation images. If you take them using the same poses, with the same camera and lighting within a few hours of each other then this method will work extremely well [1]. I really doubt it generalizes to the case of using it with a laptop webcam in any location with different lighting.
But maybe I am wrong, maybe there are enough bits of information in a randomly lit image of a face.
Has IBM built a product around this? I don’t know of one.
Or is the research for patent purposes only?
Someone at the University of Haifa has a sense of humor
The minisketch library I worked on can be used for near optimal (in the sense of information leak) error correction for "set like" features:
https://github.com/sipa/minisketch/
Our application is for communications efficient set reconciliation to convert Bitcoin's quadratic-overhead transaction gossip protocol (O(txn*peers)) to effectively linear (O(txn)), though the primary academic work that our work was based on were concerned with fuzzy extractors for privacy preserving (and encryption key generating) biometrics.