4 ms·

I applaud the effort but in the demo I am unsure how this could be used as a CAPTCHA. The examples I saw could be trivially solved with a bot running a simple C

I applaud the effort but in the demo I am unsure how this could be used as a CAPTCHA. The examples I saw could be trivially solved with a bot running a simple CNN image classifier model. The training data is available, existing OSM building outlines and satellite data scraped from Google, and training a image classifier would be very straightforward. I am sure a bot would have a much higher success rate than a person.

phoronixrly 2y ago
It can also be solved by a captcha-solving farm, as can recaptcha, what is your point?
- alex_duf 2y ago
  Also if a captcha solving farm ends up working on a project that is benefiting the world, it's still a net improvement over clicking a box that says "I'm not a robot", and an improvement over exclusively helping google with its OCR and image recognition.
  - remram 2y ago
    The question you have to ask is how inaccurate the data can be while passing the CAPTCHA challenge. The people trying to pass it at scale using bots or farms don't care to pass it correctly.
    - alex_duf 2y ago
      I agree, but since you don't know which part of the data is for validation and which one is for training, and since you can assume such a system will only accept data that makes consensus, you tend to only accept data that is valid. That's the beauty of it.
berkes 2y ago
You sound like you know it to be trivial to build this. If you can build it, (which I doubt) please contribute to OSM by getting involved with e.g. HOT OSM. https://www.hotosm.org/get-involved https://www.hotosm.org/get-involved And I doubt it is that easy, because smart people, at a.o. HOT OSM, have been building exactly this: tools to automatically categorize, detect, rate etc, mapping data from satellite imagery. Not to bypass a CAPTCHA, but to make editing and improving OSM data easier. They then concluded that some problems there are hard to solve with AI, and e.g. need better and more training data. If it's hard to solve with AI, then building that CAPTCHA bypass bot is probably just as hard if not harder.
- teruakohatu 2y ago
  > And I doubt it is that easy, because smart people, at a.o. HOT OSM, have been building exactly this: tools to automatically categorize, detect, rate etc, mapping data from satellite imagery. Not to bypass a CAPTCHA, but to make editing and improving OSM data easier. I apologise if I caused any offence. I certainly didn’t intend to. I claimed building a bot to bypass the captcha was trivial. For a captcha to be bypassed, it needs only perform better than a human doing the same task on average. Because humans make mistakes captchas also require a margin of error. A bypass bot does not need to be able to annotate all buildings on the earth perfectly. OSMs problem is that not that ML cannot do solve problem but rather they would like the result to be close to perfect and apply it on scale to the entire earth. This is a harder problem than detecting errors in a handful of images. There is a lot of academic literature on the subject of “building footprint extraction” if you are interested: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=building+footprint+extraction&oq=building+footprint https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=buil...

phoronixrly 2y ago

It can also be solved by a captcha-solving farm, as can recaptcha, what is your point?

alex_duf 2y ago
Also if a captcha solving farm ends up working on a project that is benefiting the world, it's still a net improvement over clicking a box that says "I'm not a robot", and an improvement over exclusively helping google with its OCR and image recognition.
- remram 2y ago
  The question you have to ask is how inaccurate the data can be while passing the CAPTCHA challenge. The people trying to pass it at scale using bots or farms don't care to pass it correctly.
  - alex_duf 2y ago
    I agree, but since you don't know which part of the data is for validation and which one is for training, and since you can assume such a system will only accept data that makes consensus, you tend to only accept data that is valid. That's the beauty of it.

berkes 2y ago

You sound like you know it to be trivial to build this. If you can build it, (which I doubt) please contribute to OSM by getting involved with e.g. HOT OSM. https://www.hotosm.org/get-involved https://www.hotosm.org/get-involved And I doubt it is that easy, because smart people, at a.o. HOT OSM, have been building exactly this: tools to automatically categorize, detect, rate etc, mapping data from satellite imagery. Not to bypass a CAPTCHA, but to make editing and improving OSM data easier. They then concluded that some problems there are hard to solve with AI, and e.g. need better and more training data. If it's hard to solve with AI, then building that CAPTCHA bypass bot is probably just as hard if not harder.

teruakohatu 2y ago
> And I doubt it is that easy, because smart people, at a.o. HOT OSM, have been building exactly this: tools to automatically categorize, detect, rate etc, mapping data from satellite imagery. Not to bypass a CAPTCHA, but to make editing and improving OSM data easier. I apologise if I caused any offence. I certainly didn’t intend to. I claimed building a bot to bypass the captcha was trivial. For a captcha to be bypassed, it needs only perform better than a human doing the same task on average. Because humans make mistakes captchas also require a margin of error. A bypass bot does not need to be able to annotate all buildings on the earth perfectly. OSMs problem is that not that ML cannot do solve problem but rather they would like the result to be close to perfect and apply it on scale to the entire earth. This is a harder problem than detecting errors in a handful of images. There is a lot of academic literature on the subject of “building footprint extraction” if you are interested: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=building+footprint+extraction&oq=building+footprint https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=buil...