more like 2-3 hours.
it's innovative for sure, but any well-trained convolutional neural network will have no problems with it.
hell, even a poorly trained one with off-the-shelf software had no problem:
![]()
![]()
samples were horses, trucks, motorcycles, joggers, and fish -- all with less than 10 training images per category.
no doubt could get 80%+ certainty, if not higher, with more time/images fed into it.
high-accuracy cracking of these captchas is pretty much simple as capturing a few frames from the animation for the test set and running that against (named, for output) trained sets.
always fun to see new captcha approaches though.
one of the better (read: no boot on neck) ways to fuel the AI arms race.
Numenta ftw