How to use the generated PAGE-XML as input to TrOCR?

[jarrod-dexter]

Thanks for the great library!

I am new to CV/NLP and I am trying to figure out what to do with the generated PAGE-XML (I struggled to find accessible online content about the format).

I am essentially trying to figure out how to extract a set of single sentences I could use as input to TrOCR as mentioned in the following post: https://github.com/microsoft/unilm/issues/451#issuecomment-961408406.

Would it be possible to provide some guidance?

Many thanks!

J

[cneud]

Sorry for the late reply!

Here is some background info on the PAGE-XML format:

• https://github.com/PRImA-Research-Lab/PAGE-XML - repo with schema, some docs and examples
• https://ocr-d.de/en/gt-guidelines/trans/trPage.html - explore the schema
• https://primaresearch.org/publications/ICPR2010_Pletschacher_PAGE - the original publication introducing the format
• https://github.com/PRImA-Research-Lab/prima-core-libs - Java library for working with PAGE-XML from its developers
• https://github.com/OCR-D/core/tree/master/ocrd_models - Python helpers for working with PAGE-XML (in OCR-D context)

More specifically about your question, I am not too familiar yet with TrOCR, but I assume you would have to extract the text lines (TextLine) with their bounding boxes/polygons (Coords) from the PAGE-XML output of Eynollah to derive text line images and feed the according snippets to TrOCR for text recognition/prediction.