I am not asking about facial recognition that people use as a sort of password. I mean the industrial-strength stuff the police use.
How much computing horsepower does such a system take? Might it soon be possible for me to subscribe to a system like the police have?
Do you have an example of what “industrial strength” is?
The Microsoft facial recognition requires special cameras (infared and other technology) that can’t be spoofed by holding up a photo. Surface Pro is pretty robust with an Intel Core M, so it doesn’t require a ton of horsepower.
Would the Surface Pro be non-industrial or industrial in your example?
I am not well-educated on the subject. May I presume a police computer works by sending some sort of numerical depiction of the face to a vast database in The Cloud? If so I suppose the issue would be subscribing to the database.
Can you tell us what you intend to do more precisely? Scan live camera feeds to find your ex? Sort your family pictures?
Google Photos will group your collection of pictures according to who’s in them. Even if the people are wearing sunglasses or the pictures are somewhat fuzzy. I haven’t seen a way to do this kind of recognition outside your own account (it’s surely possible, but not offered).
According to articles, Facebook also has the technical abilityto find all pictures of an individual in everybody’s accounts, or conversely to give names to the people in a group picture, but I haven’t seen those offered as a service.
I suppose what I really want for myself is a Google Glass thing to help me with names. But as a more practical matter, could a hobbyist set up a camera someplace and identify people as they walk by? Would such a system be doable at a reasonable price?
I tinker with Home Assistant, an open source home automation platform and it appears the have a few plugins that do it: Integrations - Home Assistant
I have no idea how well they work, since my stuff it mostly flipping switches at this point, but it does seem like a hobbyist could do it.
There’s a wide range of what you might mean by “facial recognition”.
To recognize your face and only your face is something a laptop can be set up to do by a knowledgeable person. But it might be spoofable if you’re not really careful.
To take a shot off a laptop camera and find out who that person is from a large pool of millions of people is something else entirely.
Since you mention “like the police have” you seem to be tending towards the latter.
First you need to already have a large database of IDed photos. And more than one per person preferably. Then you need a lot of computing power. A lot.
The first is easily solved with a lot of money. There are companies out there that will sell you such data. Usually obtained in iffy ways via Facebook and such type deals.
The second is mostly solved by buying a lot of computing power. E.g., via Amazon’s cloud computing service.
The rest comes down to having the software. AFAIK, Google has AI engines that can be used for this purpose.
Might we expect (hope, fear) that the required data and computing power will come to ordinary civilians in the next dozen years?
There’s no issue with computing power. Your phone could do it.
It’s access to a deep enough database to make it useful that’s the issue. Hopefully, that will never happen.
??? To ID one given person, for example, a good phone could do it. To find a match from from a large number of people (millions or more), it’s going to take a while, a lot of bandwidth, etc.
Microsoft’s image collection software running on my desktop several years ago would let me tag people in photos and then line up other pics of them as suggestions. Is this one Michael? How about this one?
It was pretty good, and ran on my computer. I’ve been hesitant to let google do the same even though they now have all my photos.
The compute power is there. Amazon Web Services, Microsoft Azure, etc. Basically, individuals can set up an account with a credit card and have as much computing power in the cloud as you need. The providers just charge by the byte.
Not sure about a database of photos for individuals.
The computing power required depends on your accuracy needs and how big your database is. If you have 50 friends you want to identify in their frontal selfies and a 80% accuracy rate is good enough, it takes trivial computing power. iPhoto, Google Photos, Facebook, and Lightroom can all do that for you.
If you have a fleet of hunter-killer drones that assassinate individuals on sight from different angles but only when they match with a 95%+ certainty, out of a world population of 8 billion people, that is harder for a variety of reasons.
A Google Glass type solution is already viable – all the tech is definitely there – but privacy concerns stop it from being made widely available (at least that was Google’s reasoning for banning it from Glass). This isn’t really a tech problem anymore (not at the scale you’re talking about) but a social or profitability one.
I worked several years ago for a large retail firm that was considering setting up a system to exclude shoplifters.
The system would monitor the cameras at the store entrances, check the faces against a database of previously caught shoplifters (from any of their thousands of stores), and notify store security when there was a tentative match. Security would then approach the person, and request ID to verify the identification. If it was that shoplifter (or they refused to show ID) they would be either escorted out of the store, or delivered to the police charged with trespassing (and possibly, violating the terms of their probation).
This was several years ago, and at the time, it was questionable if the technology would support this – at a reasonable cost; costs were very important to this company.
Now, I’m pretty sure this technology could be done, and at a reasonable cost. There are several advantages that make it workable:
I don’t see any technical reason such a system couldn’t be implemented now. It may have already been done, without announcing it. (But I’d think a store would announce it, just to scare shoplifters away.) I know some high-end businesses have a system that recognized good customers, so that sales staff can greet them by name.
You might be surprised. These algorithms aren’t matching people’s faces with a pixel-by-pixel photo comparison or anything like that. They determine a bunch of metrics about faces (depth, angle, and distance between various features) with a fair degree of precision, and use that to make a matrix of values that “fingerprint” the person.
The actual matrix is fairly small; a couple hundred bytes of data, even uncompressed, would be enough to identify a person to a high degree of accuracy.
You could store the ID info of a hundred million people (more than are actually in such databases, today) in less than 20 gigabytes (a lot less, since the data is highly compressible) – easily enough to fit on a modern smarphone, and search it in a fraction of a second.
Haven’t the casinos in Las Vegas been doing this for years?