Research map / Image generation / Autoregressive and latent generation
M5HisDoc: A Large-scale Multi-style Chinese Historical Document Analysis Benchmark
Yongxin Shi, Chongyu Liu, Dezhi Peng, Cheng Jian, Jiarong Huang, Lianwen Jin
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 |
|---|---|---|---|
| raise IoU threshold from 0.5 to 0.7 | DBNet++ · M5HisDoc-H text line detection | F1-score: 90.77 -> 38.57 | Hurt |
| raise IoU threshold from 0.5 to 0.7 | PAN · M5HisDoc-H text line detection | F1-score: 86.22 -> 71.55 | Hurt |
| raise IoU threshold from 0.5 to 0.7 | TextSnake · M5HisDoc-H text line detection | F1-score: 90.85 -> 84.76 | Mixed |
| raise IoU threshold from 0.5 to 0.7 | YOLOX · M5HisDoc-H character detection | F1-score: 97.35 -> 84.76 | Hurt |
| pass detected text lines through CRNN for end-to-end evaluation | Mask R-CNN · M5HisDoc-H | F1-score vs 1-NED: 84.36 -> 66.17 | Hurt |
| Cascade R-CNN instead of Mask R-CNN | regression-based detector · M5HisDoc-H text line detection IoU 0.5 | F1-score: 84.36 -> 86.74 | Improved |
| CTC-based CRNN instead of attention-based ASTER | text line recognizer · M5HisDoc-H | correct rate: 86.95 -> 90.22 | Improved |
| train with character bounding box annotations (Peng et al.*) (ablation) | Peng et al. · M5HisDoc-H text line recognition | correct rate: 88.16 -> 89.97 | Improved |
| heuristic reading order instead of Augmented XY Cut | reading order prediction · M5HisDoc-H | ARD: 14.63 -> 5.27 | Improved |
| evaluate on distorted M5HisDoc-H instead of M5HisDoc-R | LayoutReader · reading order prediction | ARD: 5.49 -> 8.68 | Hurt |
| ViT backbone instead of SwinTransformer | character recognizer · M5HisDoc-H | top-1 accuracy: 94.68 -> 94.09 | Hurt |
| measure macro accuracy instead of top-1 accuracy | ResNet50 · M5HisDoc-H character recognition | accuracy: 94.40 -> 69.61 | Hurt |
Closest papers by meaning
- Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents (AAAI 2024 · similarity 0.89)
- MosaicDoc: A Large-Scale Bilingual Benchmark for Visually Rich Document Understanding (AAAI 2026 · similarity 0.86)
- MPDocBench-Parse: Benchmarking Practical Multi-page Document Parsing (NeurIPS 2026 · similarity 0.85)
- Reference-Based Post-OCR Processing with LLM for Precise Diacritic Text in Historical Document Recognition (AAAI 2025 · similarity 0.84)
- Predicting the Original Appearance of Damaged Historical Documents (AAAI 2025 · similarity 0.83)
- M2Doc: A Multi-Modal Fusion Approach for Document Layout Analysis (AAAI 2024 · similarity 0.83)
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.
