Comment by peteretep | Hacker News Reader

Comment by peteretep | Hacker News Reader

So all through this I’m thinking “just OCR it and be done”, and we get to:

> Why not OCR all the time? > Running OCR on a PDF scan usually takes at least an order of magnitude longer than extracting the text directly from the PDF.

... so? Google can OCR video and translate it in something that feels like real-time; what PDF processing are they doing that is so performance bound?

> Difficulties with non-standard characters and glyphs OCR algorithms have a hard time dealing with novel characters, such as smiley faces, stars/circles/squares (used in bullet point lists), superscripts, complex mathematical symbols etc.

Sure, but more than the random shit you find in PDFs anyway?

> Extracting text from images offers no such hints

Finding an algorithm that approximates how a human approaches a page layout doesn’t feel like it would be all that hard.

Obviously it’s very easy to stand on the sidelines and throw stones, but parsing PDFs using anything other than OCR + some machine learning models to work out what the type of a piece of text feels like pretending we are still constrained by the processing costs of 5 years ago

> Google can

1) Do you have trillion or so dollars at your beck and call? If not, you're not Google.

> Finding an algorithm that approximates how a human...

2) ...is generally nigh impossible even for someone with Google's resources (e.g. Waymo, although when it comes to reading, it's somewhat usable). Also, look at 1)

Unless by approximate you mean toddler level. In that case:

3) The approximation is probably useless

> Finding an algorithm that approximates how a human approaches a page layout doesn’t feel like it would be all that hard.

"In CS, it can be hard to explain the difference between the easy and the virtually impossible."

https://xkcd.com/1425/

Totally agree. I worked on a project that had to try and extract tables from PDFs. It is much harder that it would first appear.

Detecting where tables are is still an active research areas. Once we know where on the page, it’s easier to parse out their structure.

To be fair, that particular comic's 5 year task was solved shortly after it was posted, at least in part to prove it was wrong.

Without having heard of or tested the solution, I'll bet anyone $1M that I can produce an image that produces an incorrect answer. Which would mean it's not "solved".

If I can produce an image that you incorrectly label as Bird or No Bird, does that mean it's accurate to say you cannot tell me if pictures have birds in them? Or is that needlessly pedantic beyond any practical use case and clearly the intended context?

No, I don't think it's pedantic and impractical. There are many use cases where the possibility of adversarial input makes heuristics a non-starter.

Please tell me the practical concerns regarding adversarial inputs against a phone app that tries to tag pictures of birds in national parks.

Well of course if you define it narrowly that's true, but computer vision has many more safety-critical applications, most obviously self-driving.

"Doing better than me", or any other human, wasn't the problem proposed. Anything other than 100% accuracy means the problem isn't solved as there will always be room for a better solution.

The problem posed was "can my app tag photos with birds". The answer is now yes for all practical purposes.

You italicized the word 'solved' to emphasize that under technical scrutiny it is not true in response to a claim about being pedantic.

100% accuracy on identifying birds is, epistemologically, impossible.

Reply on news.ycombinator.com