Hacker News new | ask | show | jobs
Show HN: Pytorch Text Recognition Tool (github.com)
41 points by s3nh_ 2484 days ago
4 comments

In the example on the readme, why does it reverse the order of "casteli" (3) and "castle" (4)? It's a bit surprising that it understand the rest of the ordering (including the "and" in the center), but flips those two.

Also, if I were a developer trying to use this, I'd be constantly annoyed at receiving a dict with keys like "0", "1", "2" rather than just getting a list.

In the example, text is not sorted by it's corrdinates but by appearence of boxes in first network. It is visible in more complex documents, that crnn network did not create boxes in descending order (word-by-word).

also, good point about the list. dictionary keys has no logical usage in this one.

This is awesome, definitely a ton of use cases for this. It would be interesting if you put some background into why you made this project in your README. Some inspiration always helps.

Also some examples of where else you've seen it applied could spark peoples imagination to help people get some more usage out of your work.

Hi, thanks for feedback! I'll add more general information. In my opinion theres a lot to do in complex document classification, I'll try to add some demo to make things more intuitive. thanks!
How to train this model in other languages?
the hardest part in training model in foreign languages is to get correctly labeled dataset. I worked with pretrain model on Polish language documents and based on this experience it is relatively good if you are using some text similarity measures. There are some examples/pretrain models with Korean/English/French language
And CRAFT stands for what?
The topic list gives the answer https://github.com/topics/craft

Fist repo: Character Region Awareness for Text Detection (CRAFT)

https://github.com/clovaai/CRAFT-pytorch

Which has a nice video demo.

https://youtu.be/HI8MzpY8KMI