OMR Model Training

Goal: train high-accuracy detectors for optical music recognition (dense, tiny symbols).

Planned model families:

• Deformable DETR
• RT-DETRv2 / RT-DETRv3
• Cascade R-CNN + FPN

Next steps:

1. Add dataset pointers and tiling/preprocessing scripts for high-res scores.
2. Create training configs for each model family.
3. Run smoke tests, then full training/eval; track metrics and checkpoints here.