What am I doing wrong?

[johnlockejrr]

I'm trying since some time to fine tune a segmentation model for Syriac script with vowels (above and below line), I'm getting closer but not enough. Kraken seems to refuse to comply :)

Ground truth (page-xml):

After fine tuning with: ketos segtrain -d cuda:0 -f page -t output.txt -q early --min-epochs 100 -cl --threads 10 --resize both --schedule reduceonplateau -i BiblIAlong02_se3_2_tl.mlmodel -o out/syrnt_cl_v1

Any idea? Should I just use bounding boxes around the baselines and not polygons?

[johnlockejrr]

Fine tuning on blla way better for text lines but loss in classes...

(ketos segtrain -d cuda:0 -f page -t output.txt -q early --min-epochs 60 --threads 10 --resize both --schedule reduceonplateau -i blla.mlmodel -o out_blla/syrnt_blla_v1)

[johnlockejrr]

Now is way better... anyway, should I add more padding?

[ketos segtrain -d cuda:0 -f page -t output-syrnt_cl_two.txt -q early --min-epochs 50 --threads 10 --resize both --schedule reduceonplateau -i blla.mlmodel -o out_blla/syrnt_two_blla_v1]