Which made me wonder why (1) it would reject invalid answers and (2) it would confuse things no human would, eg "Bus bus bus… no that’s a van, that’s clearly a van, it has Bob’s Plumbing written on it… it won’t let me get past without clicking on the van sigh.
I mean, if the aim is to train an AI, why are you ignoring the human’s answers? How do you say “No this isn’t a f—ing bus you idiot” to the captcha system? I never saw anything allowing us to do that.
Not quite. When I used to care and kind of tried to distort the training data, I would always select one additional picture that did not contain the desired object, and my answer would usually be accepted. I.e. they were aware that the images weren’t 100% lined up with the labels in their database, so they’d give some leeway to the users, letting them correct those potential mistakes and smooth out the data.
it won’t let me get past without clicking on the van
That’s your assumption. Had you not clicked on the van, maybe it would’ve let you through anyway, it’s not necessarily that strict. Or it would just give you a new captcha to solve. Either way, if your answer did not line up with what the system expected (your assumption being that they had already classified it as a bus) it would call attention to the image. So, they might send it over to a real human to check what it really is, or put it into some different combination with other vehicles to filter it out and reclassify.
Not quite. When I used to care and kind of tried to distort the training data, I would always select one additional picture that did not contain the desired object, and my answer would usually be accepted
Yes, that’s true.
That’s your assumption. Had you not clicked on the van, maybe it would’ve let you through anyway
Perhaps you should ask yourself why I wrote “it won’t let me get past without clicking on the van” rather than “It probably won’t let me get past without clicking on the van.”
I was reporting what happened, not some wild guess I made without testing.
Which made me wonder why (1) it would reject invalid answers and (2) it would confuse things no human would, eg "Bus bus bus… no that’s a van, that’s clearly a van, it has Bob’s Plumbing written on it… it won’t let me get past without clicking on the van sigh.
I mean, if the aim is to train an AI, why are you ignoring the human’s answers? How do you say “No this isn’t a f—ing bus you idiot” to the captcha system? I never saw anything allowing us to do that.
Pretty sure I’ve had “click all bicycles”, with a bicycle drawing on the road.
The first captcha they already knew the answer to. The second captcha was to build the database.
Not quite. When I used to care and kind of tried to distort the training data, I would always select one additional picture that did not contain the desired object, and my answer would usually be accepted. I.e. they were aware that the images weren’t 100% lined up with the labels in their database, so they’d give some leeway to the users, letting them correct those potential mistakes and smooth out the data.
That’s your assumption. Had you not clicked on the van, maybe it would’ve let you through anyway, it’s not necessarily that strict. Or it would just give you a new captcha to solve. Either way, if your answer did not line up with what the system expected (your assumption being that they had already classified it as a bus) it would call attention to the image. So, they might send it over to a real human to check what it really is, or put it into some different combination with other vehicles to filter it out and reclassify.
Yes, that’s true.
Perhaps you should ask yourself why I wrote “it won’t let me get past without clicking on the van” rather than “It probably won’t let me get past without clicking on the van.”
I was reporting what happened, not some wild guess I made without testing.