And now have that work without depth data, increase the population size from one person to 1+ billion, apply to moving video, aim for high precision and high recall, keep it battery friendly as you distribute to all police, somehow get the classifier of 1B unique people onto all these embedded devices, keep it up to date as people age, etc....
The article mentions that the glasses are connected to a smartphone like device. That can handle the battery and computing, all you need on the glasses portion is the camera. You don't load all the data onto the device, you just ping back to a central database (that way you also get to track every location that someone shows up at). To keep things up to date you simply pull the latest images of the person from a commercial partner who will identify and photograph them much more frequently than the government.
>Banks are beginning to use facial recognition instead of cards at cash machines while the travel and leisure industry also sees opportunities — China Southern Airlines this year began doing away with boarding passes in favour of the scheme.
If your facial recognition data in the system is out of date you will eventually find it difficult to withdraw cash or travel. People will update it themselves.
Wouldn't they do something like take all wanteds from each local precinct, up to 10000 from each village, top 10000 from local township, top 10000 from local county, top 10000 from local prefecture, etc...?
I mean you're not trying to ID every person in front of you. You want to be able to pick out known bad elements within your coverage zone.
I don’t disagree that doing image classification at scale with the parameters you’re describing is non-trivial.
But I was attempting to rebut GP’s point that a small camera is incapable of performing facial recognition. That’s plain incorrect. Speed/scale are orthogonal issues.
First, faceid is a depth camera not a photo camera. It’s also quite high resolution.
Second of course you CAN do recognition on something that’s low resolution. But if the question is how many labels can you correctly identify with high precision and recall... well... that’s a harder problem.
It’s trivial to train a CNN image classifier that can identify an individual, even when they are wearing a hat, grew facial hair, have a scar, etc