Which OCR did you use?
#5
by
iggyigor
- opened
The quality of the images is quiet bad, verified this with multipile OCR (including the built int tesseract), so i suspect that's why the model is extremely poor even on publaynet dataset.
Correct?