5 ms·

What is more interesting about that statement is how did he learn that fact? Possibly some help from the inside? Or some other method...?

Omniusaspirer 10y ago
Probably just started saving them. The value on doing so was high enough that even if it was 10's of millions it would be worth doing. It's not particularly expensive (other than time) to do so.
AznHisoka 10y ago
Just scrape.. store the captchas, use Amazon Turk, have them store the image + word in a database to recognize automatically later, and then realize the database never seems to increase in size.
striking 10y ago
He might've tried to write a computer vision solution to solve TicketMaster captchas. In order to test your solution, you need a bank of images to test from. So you download the first, say, 50k. And realize that you're getting a whole lot of repeats. 30k is still really annoying to get around without computer vision. But OpenCV is really, really fun and easy to use.
- acchow 10y ago
  It's a CV arms race - the CAPTCHAs are meant to stay ahead of state-of-the-art CV tech. That's the entire premise behind reCAPTCHA - if you've solved it, then we've sold useful real-world problems and can move onto the next ones.
  - dsp1234 10y ago
    That's the entire premise behind reCAPTCHA - if you've solved it, then we've sold useful real-world problems and can move onto the next ones. The original premise of reCAPTCHA was to help crowdsource the digitization of books by providing spam bot prevention. It's slogan was "Stop Spam, Read Books." up until 2014.
    - acchow 10y ago
      Exactly. I had a typo where I meant "solved" but put "sold". reCAPTCHA directed the efforts in the spambots arms race towards a real-world problem - reading old books.
adekok 10y ago
The birthday paradox. After a few hundred to low thousands of manual captcha's, someone goes Hey... I've seen this one before!. After 2-3 duplicates, they're pretty sure that the corpus is small. At that point, it's worth writing a script to grab as many captchas as possible. They can then manually analyze them. A back of the envelope calculation shows that with 30 people, they should be able to solve all of the captchas in a working day, even accounting for breaks.
allenz 10y ago
The files could have been numbered consecutively.
1024core 10y ago
If they had just scraped ~205 images, they would have a > 50% chance of getting a repeat.
- greeneggs 10y ago
  Indeed, by the "birthday paradox." And, by the "coupon collector's problem," the expected time to have seen all 30,000 images is about n*log(n) ~= 310,000 random images. https://en.wikipedia.org/wiki/Coupon_collector's_problem https://en.wikipedia.org/wiki/Coupon_collector's_problem
Spivak 10y ago
One method is to use capture-recapture sampling. He takes a sample of the images of size N and remembers them. Then he takes another sample of the images with size M and notes that K of them are repeated. Then an estimator for the total number of images is given by NM/K. Or more concretely, if you take a random sample of 100 images and remember them, then take another sample of 100 images and you recognize half 50 of them. Then you're reasonably confident that 100 is half of the images. (This does assume that the images presented follow a discrete uniform distribution but it's probably a good assumption given how people write software.)

Omniusaspirer 10y ago

Probably just started saving them. The value on doing so was high enough that even if it was 10's of millions it would be worth doing. It's not particularly expensive (other than time) to do so.

AznHisoka 10y ago

Just scrape.. store the captchas, use Amazon Turk, have them store the image + word in a database to recognize automatically later, and then realize the database never seems to increase in size.

striking 10y ago

He might've tried to write a computer vision solution to solve TicketMaster captchas. In order to test your solution, you need a bank of images to test from. So you download the first, say, 50k. And realize that you're getting a whole lot of repeats. 30k is still really annoying to get around without computer vision. But OpenCV is really, really fun and easy to use.

acchow 10y ago
It's a CV arms race - the CAPTCHAs are meant to stay ahead of state-of-the-art CV tech. That's the entire premise behind reCAPTCHA - if you've solved it, then we've sold useful real-world problems and can move onto the next ones.
- dsp1234 10y ago
  That's the entire premise behind reCAPTCHA - if you've solved it, then we've sold useful real-world problems and can move onto the next ones. The original premise of reCAPTCHA was to help crowdsource the digitization of books by providing spam bot prevention. It's slogan was "Stop Spam, Read Books." up until 2014.
  - acchow 10y ago
    Exactly. I had a typo where I meant "solved" but put "sold". reCAPTCHA directed the efforts in the spambots arms race towards a real-world problem - reading old books.

adekok 10y ago

The birthday paradox. After a few hundred to low thousands of manual captcha's, someone goes Hey... I've seen this one before!. After 2-3 duplicates, they're pretty sure that the corpus is small. At that point, it's worth writing a script to grab as many captchas as possible. They can then manually analyze them. A back of the envelope calculation shows that with 30 people, they should be able to solve all of the captchas in a working day, even accounting for breaks.

allenz 10y ago

The files could have been numbered consecutively.

1024core 10y ago

If they had just scraped ~205 images, they would have a > 50% chance of getting a repeat.

greeneggs 10y ago
Indeed, by the "birthday paradox." And, by the "coupon collector's problem," the expected time to have seen all 30,000 images is about n*log(n) ~= 310,000 random images. https://en.wikipedia.org/wiki/Coupon_collector's_problem https://en.wikipedia.org/wiki/Coupon_collector's_problem

Spivak 10y ago

One method is to use capture-recapture sampling. He takes a sample of the images of size N and remembers them. Then he takes another sample of the images with size M and notes that K of them are repeated. Then an estimator for the total number of images is given by NM/K. Or more concretely, if you take a random sample of 100 images and remember them, then take another sample of 100 images and you recognize half 50 of them. Then you're reasonably confident that 100 is half of the images. (This does assume that the images presented follow a discrete uniform distribution but it's probably a good assumption given how people write software.)