[zmnd]

We have tried neural nets before as well. But technologies and constraints are changing. I just took a screenshot of the Facebook feed, and segment-anything model was able to identify most of the elements, even without proper fine-tuning. Scrapping billions of elements with classifications is also something fairly trivial.

I, perhaps naively, think that training a better model(s) that helps with accessibility is a much better path forward for users that need it. Just a genuine question, not sure why I am getting downvoted. By no means, I am implying that this is not important.