Training or finetuning segmentation/recognition model on new data

[johnlockejrr]

I want to train a new model for segmentation then recognition on a pretty decent size of a ground-truth, about 730 pages of 24 different manuscripts. Is the same script by many hands (Hebrew/Aramaic Samaritan to be precise). Should I finetune some models or should I start from scratch? So far, with other data (same script) I have pretty good results finetuning but the data was way smaller 50/100/200 pages. Thanks!

[johnlockejrr]

And, by the way, should I train with topline, centerline or baseline? I think in this case centerline would fit, just want to be sure.

[johnlockejrr]

Just a fun fact:

• Issue Training or finetuning segmentation/recognition model on new data #666
• I was segmenting the 666th page of the ground truth
  ROFL :)

[mittagessen]

For the segmenter I'd start with fine-tuning the base model that you can find in the kraken repository. For the recognizer you might try to train a generalized model based on BiblIA instead of training from scratch but you'll still easily get a good quality recognition model when training from scratch.

And, by the way, should I train with `topline`, `centerline` or `baseline`? I think in this case `centerline` would fit, just want to be sure.

The topline, centerline, baseline switches are hints for the polygonizer during inference. The training procedure isn't affected by them. Fundamentally, the switches tell the polygonizer how you annotated the ground truth so it can internally translate the baseline slightly upwards, topline slightly downwards, and centerline not at all to make polygonization more robust. It looks like you annotated centerlines and the polygons already look good so the `-cl` switch is most likely appropriate.

[johnlockejrr]

Thank you so much @mittagessen