Slashdot Mirror
Analyzing (All of) Star Trek With Face Recognition
An anonymous reader writes "Accurate face recognition is coming. Pittsburgh Pattern Recognition, a face recognition start-up spun out from Carnegie Mellon University, has posted a tech demo showing an analysis of the entire original Star Trek series using face recognition. The online visualization includes various annotated clips of the series with clickable thumbnails of each character's appearance. They also have a separate page showing the full data of all the prominent characters in every episode including extracting thumbnails of each appearance." Their software can recognize frontal or near-frontal face instances.
know the name of that red shirted guy?
Unless you're looking pretty much straight-on towards the camera, this software doesn't appear to work. It does appear to be able to track a face over multiple frames if it can recognize it in one frame, but if you have 30 seconds where no suitable frame occurs, the software doesn't know who it is, even if it's pretty blatantly obvious to a human who it is.
paintball
We forehead-challenged beings demand you stop your software discrimination!
About a year or two ago
What I wanted was for face recognition software to become more general so you could search for movies using vague memories from your childhood:
"Girl on boat", "Wheat field", "Yellow flag"
With an advanced enough search engine, you could tag everything automatically.
I didn't think of privacy concerns though, I guess thats a good point.
Are you serious? Not only is it a fantastic demonstration of face recognition technology under real world conditions, but it's also incredibly useful.. if you have a little vision. How many times have you been watching a tv show and said "wow, where have I seen that guy before?" To find out these days I typically do:
1. Take note of the show I'm watching and the episode name (if given).
2. Go to imdb and hope they have specific info on that specific episode.
3. Try to guess what the character's name was, and take note of the actor's name.
4. Click through to the actor's filmography.
And, most typically, one of those steps fails. Now imagine if your tivo or other media playing device had face recognition technology like this. You'd just press one button and it would put boxes around all the faces on-screen, you'd select the face you're interested in and it would immediately tell you the name of the actor, the name of the character that actor is playing in this episode, how many other episodes of this series that he's in, and the actor's entire filmography. That's a real product that I'd actually buy!
How we know is more important than what we know.
Jesus, that's a lot of work to go through to figure out that Bruce Campbell has been in a shitload of B-movies.
If you let it play a few minutes you'll see it indenify Spock and then in the next scene he comes up as unknown even though he's facing the camera. The system seems to fail when he arches his eyebrows.
Checkout http://facemining.pittpatt.com/S3E75/, Scotty shows up under Kirk twice, and thats with just one try.
Or, http://facemining.pittpatt.com/S1E12/ actor 0117 has an odd match on my second peek.
They might want to try shirt matching.
(3) probably most important: out of copyright.
Really?
Could you post the journal of all the things you did in public places today? I'm interest to know. Begin by telling me all the address of all the houses and buildings you entered. I mean... you can't expect privacy as soon as you leave your door, right? So I'm sure you won't mind if I know, right?
BTW, Slashdot is certainly a public place and so hiding behind a nickname should not be expected. Could you give us your real name please?
OCR is easy, obviously, with all the CAPTCHA news going around.
Bullshit.
OCRs typically boast 99% accuracy -- which sounds good until you realize that this means an error in every line or two of text.
CAPTCHAs only need to be able to solve correctly a small percentage of the time to be effective -- even smaller, given humans can screw them up, too, and that problem is getting worse. So for example, Gmail couldn't just blacklist your IP for trying to register gmail accounts, without seeing quite a lot of abuse -- and botnets make IPs almost irrelevant anyway. But even 10% accuracy, which would result in absolutely unreadable OCR, would still mean that out of every 10 gmail accounts you attempt to sign up for, you get one fully functional account.
Which is damned good, for a spammer.
But, it's though that by about 2025 the number of transistors and speed of processors will be such as to rival the brain and after that point all bets are off. It will be an exciting 15 years in AI research.
I'll place a bet: We don't currently understand the human brain very well. How do you suppose we'll be able to emulate it? And your guess of 15 years seems very optimistic...
Put another way, if I gave you a brand-new, top-of-the-line computer -- for the sake of argument, let's say it's a fully loaded Mac Pro -- only with the hard drive completely formatted, could you make it do anything useful?
I'll make it slightly more realistic. I'll give you what Linus Torvalds had: A copy of Minix and a C compiler. And of course, you've got more hardware than he does. Could you just write a modern OS?
If you assume that the raw power will let us "evolve" an AI, I'm going to suggest that it takes much more hardware to evolve a program into being than it does to run it.
But, if we imagine a conscious program we can imagine a being who can 'image' every moment of life (or of their brain), save it, and even rewind backwards, or stop and start states, easily. If you're an AI and you see something you don't want to remember, just rewind a bit and it's gone forever :P
Yes, the last 15 years or more of science fiction -- cyberpunk, in particular -- make clear just how cool it would be for an AI to exist. That doesn't mean we're anywhere close.
human intelligences uploaded into the machine
Here's the uncomfortable truth: It may well be that we create AI, but no means to "upload" ourselves. Ever. The best we can do is create AI children.
And they might not like us very much. See the other side of cyberpunk -- distopian futures with robotic overlords. (Terminator comes to mind.)
Don't thank God, thank a doctor!