Unsatisfying results about ochre

[jvdzwaan]

I had similar results (although I only trained for a single iteration, it was clear performance was not as good as for the other data I tried). My first guess was the ICDAR 2017 data was not as clean as this other data, and this is why I started working on analyzing what kinds of OCR errors occur in different datasets and what kinds of errors can and cannot be fixed by training lstms. However, due to other obligations, I have not finished this work. I plan to work on it in the near future.

[Mohit-soni-bhagwan]

Hi,
I'm trying to test ochre on the ICDAR 2017 dataset (English only). I'm not using the workflow but operating ochre myself:

1. I took all the English monographs and periodical and clean it from special chars.
2. I split the files into 2 folders (ocr and gs) and needed.
3. I've created alignment files using Hirschberg's algorithm and write JSON files.
4. I've called: create_data_division.py
5. I've called: lstm_synced.py (bilstm, lower=false)
6. then on each file I've called: lstm_synced_correct_ocr.py

And when using ocrevalUAtion I've calculated the CER and WER and got unsatisfying results. in most cases the results as less than the ocr itself.

Am I doing something wrong?
Did you get better results?

Can you tell how did you ran it without workflow, it would be very helpful