Research map / Vision-language and multimodal / Multimodal LLMs
MEH: A Multi-Style Dataset and Toolkit for Advancing Egyptian Hieroglyph Recognition
Maksim Golyadkin, Valeria Rubanova, Aleksandr Utkov, Dmitry Nikolotov, Ilya Makarov
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add synthetic word data (diffusion-rendered) to render-only pretraining (ablation) | TrOCR Large · MEH real test set | CER (%): 61.0 -> 50.7 (before fine-tuning); 33.8 -> 32.2 (after) | Improved |
| add synthetic clause data on top of synthetic words (ablation) | TrOCR Large · MEH real test set | CER (%): 50.7 -> 46.0 (before fine-tuning); 32.2 -> 31.0 (after) | Improved |
| add real training data after synthetic pretraining (ablation) | TrOCR Large · MEH real test set | CER (%): 46.0 -> 32.1 (before fine-tuning); 31.0 -> 30.6 (after) | Improved |
| add synthetic word data (ablation) | TrOCR Base · MEH real test set | CER (%): 69.1 -> 55.1 (before fine-tuning); 40.2 -> 34.8 (after) | Improved |
| add synthetic word data (ablation) | MSB · MEH real test set | CER (%): 117.3 -> 81.0 (before fine-tuning); 79.7 -> 62.7 (after) | Improved |
| add synthetic clause data on top of synthetic words (ablation) | MSB · MEH real test set | CER (%): 81.0 -> 80.5 (before fine-tuning); 62.7 -> 62.9 (after) | No clear effect |
| add real training data to MSB pretraining (ablation) | MSB · MEH real test set | CER (%): 80.5 -> 63.2 (before fine-tuning); 62.9 -> 57.4 (after) | Mixed |
| add real data to TrOCR Small after synthetic clause pretraining (ablation) | TrOCR Small · MEH real test set | CER (%): 59.8 -> 52.9 (before fine-tuning); 42.2 -> 42.9 (after) | Mixed |
| replace default tokenizer with extended tokenizer (one token per Gardiner code) (ablation) | TrOCR Large · Clauserend | CER (%): 5.6 -> 3.0 | Improved |
| replace default tokenizer with extended tokenizer (ablation) | TrOCR Small · Wordrend | CER (%): 8.4 -> 9.0 | Mixed |
| scale TrOCR from Small to Large (ablation) | TrOCR · Printed held-out style | CER (%): 31.5 -> 18.9 | Improved |
| scale TrOCR from Small to Large (ablation) | TrOCR · Physical held-out style | CER (%): 68.1 -> 45.3 | Improved |
Closest papers by meaning
- ProtoSnap: Prototype Alignment For Cuneiform Signs (ICLR 2025 · similarity 0.85)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.85)
- Muharaf: Manuscripts of Handwritten Arabic Dataset for Cursive Text Recognition (NeurIPS 2024 · similarity 0.83)
- Towards Automated Chinese Ancient Character Restoration: A Diffusion-Based Method with a New Dataset (AAAI 2024 · similarity 0.83)
- Reference-Based Post-OCR Processing with LLM for Precise Diacritic Text in Historical Document Recognition (AAAI 2025 · similarity 0.82)
- Component-Level Segmentation for Oracle Bone Inscription Decipherment (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
