Closed frytoli closed 2 years ago
Dear Olivia,
"newspaper textline" model uses a similar model structure as Dhsegment https://github.com/dhlab-epfl/dhSegment but ours is deeper. And the encoder part is benefited from pre-computed weights. So in fact it is not fine-tuned from an older textline model. You can train your own model or fine-tune our model using https://github.com/qurator-spk/sbb_pixelwise_segmentation .
Thank you!!
Firstly, thank you for this project! It's really cool.
I see in your README that you fine-tuned a previously trained textline detection model resulting in your final "newspaper textline" model. After digging into DH Segment's references, I believe the original model came from READ-BAD/ARU-Net by Grüning et el. Correct? Given this information, could you give me some tips on how I might now be able to fine tune your textline detection model with my own dataset? Do you have any code you might be willing to share?
Thanks!