Broken/split lines in segmentation

[johnlockejrr]

I'm trying to train/finetune a model with a pretty constant dataset of one column. No matter how I train from scratch or finetune it I always end up with broken/split lines here and there, what can be the issue?

How I tryed to train the model:

ketos segtrain -d cuda:0 -f page -t sam_40_mss.txt -q early --min-epochs 50 --resize both --schedule reduceonplateau -i /home/incognito/kraken-train/kraken-5.2.9-py3.11/lib/python3.11/site-packages/kraken/blla.mlmodel -o sam_40_mss_seg/sam_40_mss_seg_v1
ketos segtrain -d cuda:0 -f page -t sam_40_mss.txt -q early --min-epochs 40 -cl --resize both --schedule reduceonplateau -i /home/incognito/kraken-train/kraken-5.2.9-py3.11/lib/python3.11/site-packages/kraken/blla.mlmodel -o sam_40_mss_seg/sam_40_mss_seg_v2 :
ketos segtrain -d cuda:0 -f page -t sam_40_mss.txt -q early --min-epochs 40 --resize both --schedule reduceonplateau -i BiblIAlong02_se3_2_tl.mlmodel -o sam_40_mss_seg/sam_40_mss_seg_v3
ketos segtrain -d cuda:0 -f page -t sam_40_mss.txt -q early --min-epochs 40 --resize both -tl --schedule reduceonplateau -i BiblIAlong02_se3_2_tl.mlmodel -o sam_40_mss_seg/sam_40_mss_seg_v4
ketos segtrain -d cuda:0 -f page -t sam_40_mss.txt -q early --min-epochs 40 -cl --schedule reduceonplateau -o sam_40_mss_seg/sam_40_mss_seg_v5

Example output:

Below is the training from scratch, no broken lines but the polygon masks are awful...

[johnlockejrr]

UPDATE:

I finetuned my data with ubma_segmentation.mlmodel and now seems to do the job... I'm not sure how they trained the model, I'll ask @stweil