beaunix/Thoth-Sphinx-egyptian-hieroglyphs

Thoth-Sphinx — Egyptian Hieroglyphs Multi-Sign Detection Dataset

Dataset Summary

This dataset provides multi-sign, multi-cartouche annotated images of Middle Egyptian hieroglyphic inscriptions for object detection, in YOLO format. To our knowledge, no other publicly available dataset combines:

• Multiple signs per image (average ~34 instances/image), rather than isolated single-glyph crops
• Royal cartouche detection as its own class, with signs annotated inside the cartouche boundary
• 150 Gardiner sign classes, curated and refined across nine training iterations (v1 through v9)

Existing public resources (e.g. the Kaggle GlyphDataset) provide single-glyph, single-sign-per-image crops, which do not reflect the real-world task of reading an inscription: locating and ordering many signs across a photographed wall, stela, or cartouche.

Dataset Structure

The dataset is split into two clearly separated subsets:

real/

Photographs of genuine hieroglyphic inscriptions (temple walls, stelae, cartouches) collected and manually annotated over three months, sourced from public archaeological photography (e.g. pyramidtextsonline.com, Flickr) and personal fieldwork references. Annotated in Roboflow.

synthetic/

Auxiliary mosaic images assembled from cropped, individually-verified Gardiner sign instances, used to balance underrepresented classes during training. These are not photographs of real inscriptions — they are synthetic compositions and are labeled as such. Users should not treat this subset as representative of real epigraphic conditions (lighting, erosion, stone texture).

We recommend that any published benchmark using this dataset report results on real/ alone, using synthetic/ only as supplementary training data, to avoid inflated metrics from synthetic composition artifacts.

Classes (150 total)

Full Gardiner sign list plus cartouche and unknown (catch-all for signs outside the 150-class vocabulary). See class_map.json for the complete index-to-name mapping.

Class transmutation history (v9)

Three class slots were relabeled rather than added, because the detection head is architecturally fixed at 150 classes. The original signs at these indices had too few training instances to ever be learnable (1-3 instances); the replacement signs were common in our corpus and photographs but previously had no dedicated class (annotated as unknown):

Index	Former class	Now represents	Rationale
30	E10 (goat)	M4 (time palm / year ideogram)	E10 had ~7 instances; M4 is frequent in funerary/religious texts
61	I15 (coiled serpent)	F34 (heart)	I15 had ~3 instances; F34 is a common determinative
70	M20 (papyrus thicket)	O29 (wooden column)	M20 had ~1 instance; O29 is common in architectural inscriptions

Users retraining on top of this dataset should be aware that indices 30, 61, and 70 do not correspond to the "standard" Gardiner ordering one might assume from a naive alphabetical listing.

Annotation Methodology

• Bounding boxes drawn per individual Gardiner sign, including signs inside cartouche boundaries (not just the cartouche outline itself)
• Class labels follow the Gardiner sign list convention (lowercase except Aa series, which preserves the "Aa" category prefix, e.g. Aa1, Aa15)
• unknown is a catch-all for signs outside the 150-class vocabulary, used to avoid forcing incorrect labels
• cartouche is annotated as its own bounding box class, encompassing the full oval enclosure

Evaluation (v9 checkpoint, YOLOv11-large)

Validated on a 162-image held-out multi-glyph set (no train/val leakage):

Metric	Value
mAP50 (all classes)	0.925
mAP50-95 (all classes)	0.699
Precision	0.914
Recall	0.874
cartouche mAP50	0.920
cartouche Recall	0.796

Full per-class breakdown available in the accompanying training report. This work performed by Author during 3 months in Roboflow. Check Link Origina_Dataset

Known Limitations

• Some classes remain underrepresented (fewer than 5 instances in the real subset); per-class mAP for these should be treated as statistically unreliable, not as a reflection of true model performance on that class
• The unknown class is inherently heterogeneous and should not be treated as a learnable visual category
• Class transmutation (see above) means index-to-sign mapping differs from a naive Gardiner ordering — always consult class_map.json
• Dataset skews toward Middle Egyptian funerary and religious texts; administrative/legal text styles are underrepresented

Sources & Acknowledgments

• Real inscription photography: pyramidtextsonline.com, Flickr (public archaeological photography), personal fieldwork
• Lexical/statistical grounding for class prioritization decisions: BBAW/TLA Earlier Egyptian corpus (Thesaurus Linguae Aegyptiae), Dickson's Dictionary of Middle Egyptian
• Annotation platform: Roboflow

License

CC BY-NC 4.0 — Attribution required, non-commercial use only. See LICENSE for full text.

Citation

If you use this dataset, please cite it as:

@dataset{Thot_Sphinx_hieroglyphs_2026,
  title  = {Thot Sphinx Egyptian Hieroglyphs: Multi-Sign Detection Dataset},
  author = {[Your name]},
  year   = {2026},
  note   = {Version 9},
  url    = {https://huggingface.co/datasets/beaunix/egyptian-hieroglyphs}
}