8.11.05

Mechanical Turk

Last week I read a piece on slashdot about amazon's Mechanical Turk. Since humans still can do even some trivial tasks, like photo recognition, better than computers, they pay you to do said trivial tasks. Of course it was slashdotted within minutes, so I've only recently been able to try it myself.

The easiest way to get started is to find a group of hits, typically from large city, and try to find a given business address in a series of photos. The photos appear to be captures from a DV video taken from a moving vehicle. In a couple of the photos you can actually see the vehicle used to record in the reflection of windows. Chicago has a white Chevy Blazer.

You either match or discard all of the photos. You can return the series or "hit" as well. You get 3 cents per photo "approved". Approved doesn't necessarily seem to mean that there was a good match, as you can discard all of them if they suck.

3 cents seems like easy money. You can look at photos pretty quick, right? Unfortunately, it takes alot longer than i first figured. The photos are quite large, and since they're captures from video, they're grainy as well. Not to mention every one is street level. Sometimes, and in some cities, it's easy to match the address number with the a street address in the photo. Other times, you get pictures of parking lots.

It's still quite a slow interface, regardless of the size of photos being displayed.
I've switched to Opera, which seems to display and cache better. But before you can work on a series, you have to accept it. Then, if no one else has worked on it, you wait for it to load, sometimes again, and select an option. So for about a 20 seconds, you possibly get 3 cents. woot. Your answer is either approved or rejected, or abandoned. Since each group of pictures in the city your working in has an expiration date, it seems that if you work on hundreds of photos and their not approved in time, nobody gets the money. good to know.

So at this point, i've stopped dreaming this is a quick way to get easy money. It has proved fun though. I've been working in Chicago mostly, and since I've been there several times, it's fun to find recognizable landmarks, like my friends backyard!

Lots of room for improvement though. Several times you can follow the logic of addresses. you're looking for 123 damen ave. and you're only shown 119, 121, and the side of a truck. You know it's that corner in the last photo, but there's a ups truck parked otu front. Also, sometimes the capture is only of half of the store front. It might work if they gave you crop tools, but there's no way to specify.
Many businesses appear vacant, so you can infer their address, but who's to say the admin will notice? There should be a highlight feature, or a reasoning field.
Also, there's some common sense reasons not to pick the best photo. Sure that showcases the beautiful script sign out front, but the next photo doesn't show the adult bookstore next door.

So it's fun for now. hope I get my money. My scrolling finger is worn out.

0 Comments:

Post a Comment

<< Home