Slashdot Mirror
Computers Outperform Humans at Recognizing Faces
seven of five writes "According to the recent Face Recognition Grand Challenge, The match up of face-recognition algorithms showed that machine recognition of human individuals has improved tenfold since 2002 and a hundredfold since 1995. 'Among other advantages, 3-D facial recognition identifies individuals by exploiting distinctive features of a human face's surface--for instance, the curves of the eye sockets, nose, and chin, which are where tissue and bone are most apparent and which don't change over time. Furthermore, Phillips says, "changes in illumination have adversely affected face-recognition performance from still images. But the shape of a face isn't affected by changes in illumination." Hence, 3-D face recognition might even be used in near-dark conditions.'"
It's really annoying how much of this research never gets turned into product.. or, worse yet, it gets embedded in some proprietary piece of shit hardware instead of being released as a reusable component. I'd love to add some good facial recognition to my pet robot, but I'm not buying your watt sucking camera.
How we know is more important than what we know.
I wonder whether these scientists lose any sleep over how their research advances will contribute to the future of our societies.
I, for one, welcome our face-recognizing overlords...
"Nature bats last..."
Don't be silly.
If this were to be used for criminal identification, I'm sure that when they get a "hit" for a wanted suspect, that they're going to manually sift through the video, in order to figure out direction of travel etc.
These things aren't error proof, and never will be. A jury would also probably be more sawyed by seeing part of the footage than just having a prosecutor say "the computer said it was him."
If I were an (innocent) suspect, I'd much rather that I was tagged by a computer, since the video evidence would be available to criticize, than to be tagged by a witness to a crime, who are notorious for misidentifying people.
So in regards to your hypothetical question, no. At best this would be like a google search for faces, where an investigator would then further analyze the hits.
Right.
Recognition tasks are almost all inductive in nature, where performance on math is deductive. Human induction pretty well spanks machine induction at most of the things we take for granted - like recognizing and decoding faces, voices, speech, the sound of your walk, etc., etc., etc. The thing computers do least well is infer what bits of information are most important. We seem to excel at that.
Despite what the findings say, I stand by the faces thing. It sounds like the recognition algorithms got high-resolution 3D scans of human faces as input. Wake me when they can do as well as a human with low-resolution 2D scans.
That being said, it's great to see progress in this area. I can't wait until someone has to lop off my head and carry it with them in a plastic bag in order to break into my workplace. It's more grisly than taking a thumb, but much less likely to happen... I think...
I got my Linux laptop at System76.
I concur. Computers can outperform people under experimental conditions where the data provided to the computer is exactly what is required to perform the job at hand. Show me a computer that can recognize a person from a brief glimpse of the the back of their head, when they're walking away, on the other side of the street, and I'll agree they've got us licked on this one.
I always knew sometime in the future we'd be wearing those 80's wraparound sunglasses everywhere. That or one of those nifty 3-in-1 fake nose, mustache and glasses kits.
Yep, it can back up any claim.
For example, dinosaurs co-existed with humans.
If "better" is based on the standards of humans (fastest good enough guess) rather than machines (as correct as possible, complete & in depth), humans win.
Translation: Throw enough hardware at it, and the machines win? Whatever a computer has been successfully programmed to do, it's usually bloody fast at it. It sounds like a well parallelizable task that should scale easily for many years to come.
Live today, because you never know what tomorrow brings
Because that's not what face recognition software was made for. When people watched John Cleese, they knew they watched a celebrity and they also knew that not too many celebrities would dress up like this and do the silly walk. Only using all this extra information made people recognize John Cleese. Chances are that even in a small town you'd find quite a few people who, if dressed up and walking like that, would easily pass as "John Cleese". On the other side most actors/models would not be recognized by anyone reliably if they don't have their make up and if lighting differs from the studio where they usually have their pictures taken.
Face recognition software on the other side doesn't make those assumptions but instead focusses on identifying people from a large population of registered images, using no extra knowledge and making no assumptions. All the face recognition vendor test says is if you put up 1000000 random faces, people would misidentify more of these faces as John Cleese than modern algorithms would.
The problem becomes how do you "flash" an image at a computer. A computer has a perfect memory. So you can't compare humans and computers in this way. A computer could completely memorize millions of faces, or even all the faces in the world, given enough storage space. 6 Billion people x 1 MB (exaggeration) per picture is only 6 petabytes. It's a lot of data, but not out of reach. So if computers get good enough at recognizing faces, it could become a useful too in security. Think about the security guard sitting at the front desk of 20 story building. Do you think he could identify every person who walks through those doors. Would he know if you were just using a stolen security badge?
Anthropic principle: We see the universe the way it is because if it were different we would not be here to see it.