papadimas/trogs-greek-squeezes

TROGS Greek Squeezes

High-resolution raking-light scans of ancient Greek paper squeezes with letter-box annotations and Latin-proxy transcriptions. This package preserves the official TROGS / ICDAR 2026 train, validation, and test splits, with both lighting rotations for each squeeze.

Source: https://scholarworks.smith.edu/dds_data/18
Contest page: https://www.science.smith.edu/~nhowe/contest/trogs26.html

Contents

Split	Squeezes	Images
train	181	362
validation	22	44
test	21	42
total	224	448

Each row is one full scan for one rotation. Metadata includes base_id, rotation, image size, baseline angle, normalized transcript, raw transcript_raw, original raw_annotation, and JSON-encoded letter boxes / line specs.

transcript removes lacuna markers (*) and spaces to match the official CER scoring convention. Use transcript_raw or raw_annotation if marker placement matters.

Loading

from datasets import load_dataset

ds = load_dataset("imagefolder", data_dir="hf://datasets/papadimas/trogs-greek-squeezes/data")

For private repos, authenticate first with hf auth login or huggingface_hub.login().

Intended Use

• OCR / HTR for ancient Greek squeeze scans.
• Letter localization and transcription from shallow, low-contrast impressions.
• Experiments with preprocessing, line tiling, raking-light ensembling, and CER decoding.

Limitations

• Small dataset: 224 squeezes / 448 images.
• Domain-specific: paper squeezes, not manuscripts, papyri, printed text, or stone photos.
• Targets are Latin proxies for Greek characters, not Unicode Greek.
• Performance depends heavily on lighting direction, squeeze condition, and inscription preservation.

License and Citation

The source dataset is CC-BY 4.0. Please cite the original work:

@article{howe2025greeksqueezes,
  title = {Character Recognition for Greek Squeezes},
  author = {Howe, Nicholas R. and Chang, Feiran and Falbo, Isabella and Brown, Tajhini and Hershkowitz, Aaron},
  journal = {International Journal on Document Analysis and Recognition},
  volume = {28},
  pages = {345--356},
  year = {2025}
}