Recommendations for OCR labeling task

[heartexlabs]

Recommendations for OCR labeling task

[deppp]

@ReDeiPirati I've just checked, and you're right. Relations were not that well checked on the latest releases. I'll fix that and push with the next release and update you on that. Stay tuned

[deppp]

Hey @ReDeiPirati,

Might one solution be to include a textarea? You can also use relations if you want to connect a particular bbox region to a specific textarea input. Can something like the following work?

<View>
  <Header value="Put bbox around the text" />
  <RectangleLabels name="lbl" toName="img">
    <Label value="Label 1" />
    <Label value="Label 2" />
  </RectangleLabels>
  <Image name="img" value="$img" />
  <Header value="Transcribe each bbox" />
  <TextArea name="txt" />
</View>

Also, if you've planning to connect the ML model to help you annotate your dataset faster, it'd be easier to split this up into multiple projects. If you want to go that way depends on your use-case of course as well as the size of your dataset. Hope that helps!

[ReDeiPirati]

Hi @deppp,

You can also use relations if you want to connect a particular bbox region to a specific textarea input. Can something like the following work?

I've tried your solution, but it doesn't seem to work. The textarea is not linked to any particular bbox and I need a text area for each bbox I'll create. With relations I'm only able to connect two bounding boxes, but not the text area. Not to mention that the IDs of the relations seem to be wrong (not the same IDs of the bboxes that I want to relate).

Also, if you've planning to connect the ML model to help you annotate your dataset faster, it'd be easier to split this up into multiple projects.

Yes, I'm considering to connect a pretrained CTPN model later to help me in this process. Now, I'm focusing my attention to create the data labeling guide.

---

UPDATE

My bad, the textarea input works fine, the problem is that relating input to bbox is really bad in terms of usability and most of the time is not working properly (the IDs are not the ones expected).

[deppp]

@ReDeiPirati This is fixed now, please check the latest release.

In general, I think OCR type of tasks can be improved in terms of UI, but I've not found anything that works well really to get inspiration from. If you have any pointers or ideas on how the workflow can be improved, please let me know.

Best, Micahel