The other day I went galavanting about creating new accounts on social networking sites and other Web 2.0 sites, mostly social networking. One stands out among them as very cool, if a littl scarry, namely MyFaceID. It's an interactive demo of Betaface, which is a facial recognition product. See? Very cool! The jist of it is as follows:
- Create account (duh)
- Upload a bunch of photos of yourself and/or others
- System anylizes the photos, performs facial detection and shows you the "faces" it's found
- Tag the photos as you or whomever. Alternatively, delete the ones that are bad matches (it happens).
- Do something cool with them.
- Click on a face and see who in their database it matches. The database is of the Public profiles
and a large collection of celebrities.
- Mix several images together
"
Do something cool with them" is where things go a little off. Unfortunately you almost centainly won't match (at the time of this post) any "real" people. I can't say how many people are signed up but the public profiles amount to 34. I am not one of those 34. Presumably there are more people, people like me, signed up checking out the cool tech. The rub of it is that you'll only really match celebrities. There's nothing inherently wrong with that but since the system relies on lots of
good input the matches can be a little...
off. I promise you, I don't look at all like
Scarlett Johansson. No really, I don't. If I
did I would have a different carreer.
As for that second one, mixing faces together. That tempted me into triyng it straight away. The system only allows you to mix ten images at a time, for server sanity I assume. Okay, fine. After choosing five images of one person I'd tagged and five images of someone else I'd tagged I set it running.
Processing...processing...more processing...expired session, logged out Lesson learned, ten images equals too much
processing. It turns out that eight, in my case, worked out fine. The result? A somewhat blury face that looked like some poor bastardization of two people. Okay, just one person then. ... A somewhat blury face that looked like some poor pastardization of
one preson. Dammit! I have to say that sadly I wasn't terribly surprised. It looks like an averaging algorithm similar to stacking a bunch of images, changing their alpha levels and then flattening the layers. I haven't tried that yet but I ought to for comparison sake. In my opinion part of the problem comes down to sample set, ten images just isn't enough. In defense of the mixing system they present random images of average X vs. average Y, where X and Y are related sub components of a category (eg. Male vs Female). Those look fairly reasonable, though I'm sure that the sample is huge in comparison and takes a
long time to produce.
My general assestment? This is cool. It's a little scary at a base level of personal privacy and identity, which is I assume why there are only 34 public profiles. It needs some work (duh is bleeding edge stuff). Notably I think that it should provide options/features to:
- See the identifying points and lines that the system has determined.
- Accept some training input such as indicating that a match is bad or that an image is altogether wrong. My arm is not my face.
- Provide a "Save" feature for blended faces (even if they aren't spot on, they're still pretty cool sometimes). This will save processing later if someone wants to review and could allow for backgrounding the process in a queue to be finished later at low priority on the system.
- Provide that backgrounding system that I just mentioned.
- If the mix algorithm is as simplistic (and I hope it's not) as it appears a new morphing based solution would be interesting to see.
Still, very cool. Very cool indeed! So go check it out and be braver than me, make a public profile!
cheers!
