I hate them as much as you do, but you're wrong. Those storefront and traffic sign captchas are not useful for training ML models. If they were to be useful, they would be much more varied, like the original ones (used for OCR).
>Those storefront and traffic sign captchas are not useful for training ML models.
Not to get all tin-foil-hat, but this is going to sound like it, but if you have a car that has 9+ cameras upon it that drives in areas full of these, then maybe there would be some use for it for Google.
Bear in mind that I'm not saying that they are doing this but to dismiss it unequivocally as something that can't or wouldn't be done entirely ignores the premise that it could prove useful to other areas of their business, which might have a vested interest in such use (say, for example, if Google or it's parent company were trying to break into the self-driving car area[0]).
I would love to see some evidence (a link or something) of this. I see captchas that look like pretty good edge-detection discriminators- street lights in tree limbs, bicycles against brick, and so on.
Since they introduced the square-selecting captchas I have always assumed that they use it for identifying the user. I bet that depending on how you solve the captchas they can identify who you are if their system already has a theory of who you might be.
They're implemented this particular way to provide training data for image segmentation systems, they move the image around inside the frame which allows them to use a few people doing the challenge to create a boundary representation that can be used to train things like YOLO style ML systems
