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?
· Sign up or log in to comment
