This does not scale. It might work for a few people in a population of thousands, but beyond that false positives are going to be a real problem.
back
5 comments
I work in facial recognition. When you say it does not scale, how so? I have a system with 975K people in it, the US registered sex offenders database, and with a single quad core laptop a lookup takes about 8 seconds. With a server class 32+ core system the lookup is nearly real time. How does that not scale?
I think the OP was concerned about false positives, not processing times. If you ran, say, 10,000 faces through your system/day for a year, how many false positives would you get?
I assume he more meant accuracy at scale - if you have a large population how does the accuracy do? Do you wind up with many close samples or are things pretty good?
Pretty much all FR systems generate a list of matches, ranked from closest match on down. The size of the list is configurable. It is also industry standard NOT to use the list as an authority. Given a list of high matches (greater than 90%) a human can quickly filter out obvious false positives, and then the remaining are retained for further consideration. It is also industry standard NOT to rely on FR alone; combining FR with other measures reduces the ability for a false positive.
The issue I find with FR is people expecting it to be some super technology, gleaming significant information from a dust spec. Its not like that. And the media is playing it up with unrealistic descriptions. Remember the original DOOM and it's quality of graphics? That is where FR is now. We've got a ways before the journalist hype is close to reality. And the mature technology will not be FR, but a comprehensive multiple biometric measuring system capturing far more than someone's face.
So, if you’re looking for 1% of a total population your false positive rate before human intervention is 99%. That's not useful for automated systems.
Further, identical twins are going to show an image that would fool human verification making this questionable for many tasks even with human supervision.
PS: Almost 1% of the population has an identical twin. (% of births is lower, but you get 2 twins.)
There is only one true match. So the system will always be generating false positives. This technology is not an authority, but a filter. Yes, identical twins will both be identified if they are both in the system and no additional biometric measures are included. Identical twins still have different retina, and due to lifestyles identical twins beyond age 30 can be distinguished apart fairly easily.
What's the "industry" you speak of? This article is suggesting this be used by minimum wage (or below[1]) workers in shops, bars and restaurants. How well do you think your industries norms will be adhered to?
[1] UK tried (is trying?) a "you must work for free in what would be a minimum wage job in a high street shop" scheme. http://www.bbc.co.uk/news/uk-politics-24742499
The parent was talking about the statistics involved. Quoting lookup times is meaningless if those lookups aren't accurate. I could implement a system which always returns the first DB entry; lookups would take much less than 8 seconds, and by storing all other entries in /dev/null it has constant space usage ;)
How does that reliably work?
Because my impression from consumer products (say Picasa or Apple Photos) is, that starting with 35 to 50 tagged faces, the recommendations produce more false positives than correct ones.
I thought that Google/Apple also used Maschine Learning techniques for face recognition.
Keep in mind: 1) there will always be more false positives because there is only one true match, and 2) consumer FR "toys" are nothing like an industrial FR application. It is night and day the difference between anything a consumer can play with and what police, corporations, and security agencies use.
If this was deployed in the US, I'd be concerned about being shot by the police because I was false-IDed as a violent criminal. Because in the US, it's become protect the police at all costs, including the cost of my life.
The problem of false positives can and will simply be farmed out to a mechanical turk type of process ... a farm of low-cost workers will simply review hits, and if there's any possibility that it's a match, it will be escalated up several levels until action is taken.
It's not going to be great for having 1 robbery and nabbing the next person your cameras ID. But in this day and age were the police won't go out for burglary calls we might end up with person A identified in X surveillance system incidents over Y years. That's an interesting step forward even if it isn't damning.
Is that really a step forward? When combined with the false positives issue, there could be a lot of harassment of innocents who accidentally fit a profile.
This is clearly a play by the plastic surgery lobby to get people to have work done so they look less like known criminals.