Training Hi-SAM on custom dataset for Line-Level Text Segmentation Only

[PakizaNaeem]

Thank you for sharing the Hi-SAM models for hierarchical segmentation. I noticed you mentioned the demo model line_detection_ctw1500.pth in the Application section, which is trained on CTW1500 for text-line detection.

However, I couldn’t find any specific guidance or configuration details on how this model was trained specifically for line or word segmentation tasks. Could you please clarify:

How was the training configured to focus only on line-level segmentation?

What changes (if any) were made to datasets, annotations, or training scripts?

Can this approach be adapted for custom datasets, like Urdu text?

Any pointers or example configs would be very helpful.