Better Layout Detection Model and Handling of other elements (except text)

[Prasaderp]

Before you asking

• I have tried the PDFMathTranslate-next and give feedback in PDFMathTranslate-next

Is your feature request related to a problem?

The Doclayout model is pretty average. And thus creating a lot of structuring/formatting issues in the output PDF. I have also attached a sample. Also, this is just one sample there are many major structure issues which also includes math elements or PDF elements like lines, underscores etc..,.

Describe the solution you'd like

Can we use a better model for DocLayout detection.

Additional context

No response

[hellofinch]

Yes, we can.
Can you provide some candidate open-source models?

: )

[Prasaderp]

LayoutLMv3 - microsoft/layoutlmv3-base

How about this?

or something with better accuracy - PaddlePaddle/PP-DocLayout-L

[hellofinch]

I think this is an excellent idea!
We can try some different layout models to improve the performance.
Would you like to create a new issue in the new repo? We can discuss more with others. Maybe other maintainers and Contributors have different ideas.
: )

[Prasaderp]

Sure