I think it might be the same when they switch to other types of objects (like crosswalks or bikes). Someone's model got too good, so they had to change to something else. I also get the impression that they add delays to the tile refresh before they do that.
I suspect Google now uses robots to generate captchas for humans, under the assumption their image recognizers are far better than anyone else's. They already have some very well ones for other products (self driving cars, street view) and lots of street-level city imagery. That would explain why their captchas are so difficult for humans to solve -- they're testing if you see things like their "AI," not like other humans.
>I'd always assumed that was noise carefully tuned to throw off one machine learning model or another that was being used to beat the captcha
I was thinking it was trying to dirty up the image just like the lenses on cameras get dirty. What happens to the image recognition when there's water spots, dirt, mud, etc on the lens that keeps parts of the image obscured?
what would be the training data value in getting humans to classify images you have clean versions of, after applying simulated noise?
If you have the clean version of the image, you need to get that classified by a human - then you can throw noisy versions of it into the training set for your AI. You don’t need to ask a human, hey, I added noise to a picture of a yield sign. Is it still a picture of a yield sign?
I'd always assumed that was noise carefully tuned to throw off one machine learning model or another that was being used to beat the captcha, sort of like this: https://www.theverge.com/2017/11/2/16597276/google-ai-image-...
I think it might be the same when they switch to other types of objects (like crosswalks or bikes). Someone's model got too good, so they had to change to something else. I also get the impression that they add delays to the tile refresh before they do that.
I suspect Google now uses robots to generate captchas for humans, under the assumption their image recognizers are far better than anyone else's. They already have some very well ones for other products (self driving cars, street view) and lots of street-level city imagery. That would explain why their captchas are so difficult for humans to solve -- they're testing if you see things like their "AI," not like other humans.