SAM 2 – Photogrammetry Segmentation (Scroll & Ruler)
Fine-tuned checkpoints of SAM 2.1 (hiera-tiny), for segmenting the papyrus scroll (and, where present, a calibration ruler) in photogrammetry photos of the Herculaneum papyri. Produced by fine-tuning the prompt encoder and mask decoder of SAM 2.1 on a dataset of scroll/mask image pairs.
Code: https://github.com/ScrollPrize/villa/tree/main/foundation/sam2-photogrammetry
Checkpoints
| File | Description |
|---|---|
photo2_ruler_t_1000.torch |
For photogrammetry captures that include a calibration ruler in frame — segments both scroll and ruler. 1000 training iterations. |
photo2_t_5000.torch |
For photogrammetry captures with no ruler in frame — segments scroll only. 5000 training iterations. |
Usage
See the sam2-photogrammetry README for full setup instructions. To fetch a checkpoint directly:
wget https://huggingface.co/scrollprize/sam2-photogrammetry/resolve/main/photo2_ruler_t_1000.torch -P checkpoints/
License
Apache 2.0, matching the parent SAM 2 repository.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support