6 comments

[ 2.9 ms ] story [ 28.5 ms ] thread
I like this a lot. Couldn't see any errors in the approach. I think it is sound and good.
What's to stop a decently trained net (think like Image net level) from spamming the hell out of this?
In a strange way that would actually be great.

One of the primary goals of the project is to build a big dataset of tagged images. If a bunch of bots went through uploading and then tagging all of the images on DeCAPTCHA, that would accomplish this goal even faster!

But doesn’t that defeat the purpose of captcha? Ie, you verify someone isn’t a bot?
You bring up a totally fair point for the set of images that are on it right now.

But, as more users upload images that then must be tagged, the hope would be that eventually there is enough diversity in the dataset that being able to consistently recognize what's in the different images would require something approaching human intelligence.