microsoft / unilm

Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities
https://aka.ms/GeneralAI
MIT License
20.05k stars 2.55k forks source link

[LayoutLM] Do you plan to open source the receipt understanding task example for the SROIE dataset? #251

Open oni-on opened 4 years ago

oni-on commented 4 years ago

It'd be awesome to be able to reproduce the results obtained on the SROIE challenge.

jackie930 commented 3 years ago

feeling the same.

ruifcruz commented 3 years ago

same here. I am getting very different results from the ones presented in the paper..!

ruifcruz commented 3 years ago

Mabe it helps: https://github.com/ruifcruz/sroie-on-layoutlm

oni-on commented 3 years ago

Great work @ruifcruz ! I'm trying to run your notebook atm.

SuryaThiru commented 1 year ago

Mabe it helps: https://github.com/ruifcruz/sroie-on-layoutlm

I see, that this notebook directly uses the OCR annotations provided with the original dataset. Do we know what OCR engine was used by the original authors for the SROIE information extraction task? The LayoutLMv2 paper mentions that they "use the official OCR annotations ". Does that mean no OCR was performed and the annotations were directly used?

ruifcruz commented 1 year ago

As far as I remember (some time have passed since then), they have used tesseract (in v1). I would say that they didn't need to OCR because they already had the annotations from the contest.