Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Subjects: | |
| Online Access: | https://doi.org/10.5281/zenodo.18715381 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <h1>D-FINE Small LADaS Region Segmentation</h1> <p>A D-FINE Small region segmentation model for document layout analysis, trained on the <a href="https://github.com/DEFI-COLaF/LADaS">LADaS</a> dataset. The model detects 37 region types following the <a href="https://segmonto.github.io/">SegmOnto</a> vocabulary.</p> <h2>Architecture</h2> <p><a href="https://arxiv.org/abs/2410.13842">D-FINE</a> is a transformer-based object detector for document layout analysis. This Small variant uses:</p> <ul> <li><strong>Backbone:</strong> HGNetv2-B0</li> <li><strong>Encoder hidden dim:</strong> 256</li> <li><strong>Decoder hidden dim:</strong> 256</li> <li><strong>Transformer decoder layers:</strong> 3</li> <li><strong>Feature strides:</strong> [8, 16, 32]</li> <li><strong>Detection queries:</strong> 300</li> <li><strong>Input resolution:</strong> 1280x1280</li> <li><strong>Model size:</strong> ~42 MB</li> </ul> <h2>Training Data</h2> <p>The model was trained on the <a href="https://github.com/DEFI-COLaF/LADaS">LADaS</a> (Layout Analysis Dataset) by Thibault Clérice et al. LADaS is a multi-document diachronic layout analysis dataset comprising documents from the 17th century to the present, including monographs, PhD theses, auction catalogs, academic papers, and magazines. Annotations follow the SegmOnto vocabulary.</p> <h2>Uses</h2> <p>Intended for document region segmentation as part of an automatic text recognition pipeline with <a href="https://kraken.re">kraken</a>. Detects document zones such as main text, headings, graphics, margins, tables, and other structural elements in historical and contemporary documents.</p> <h2>Bias, Risks, and Limitations</h2> <ul> <li>The training data is predominantly French with Latin script. Performance on other languages may vary.</li> <li>Rare classes (FigureZone-FigDesc, MainZone-Maths, GraphicZone-TextualContent) have very low or zero detection scores.</li> </ul> <h2>Training Details</h2> <h3>Training Procedure and Hyperparameters</h3> <ul> <li><strong>Training regime:</strong> BF16 mixed precision</li> <li><strong>Epochs:</strong> 32</li> <li><strong>Learning rate:</strong> 1e-4</li> <li><strong>Schedule:</strong> Constant with 1000-step warmup</li> <li><strong>Weight decay:</strong> 1e-5</li> <li><strong>Batch size:</strong> 8</li> <li><strong>Image size:</strong> 1280x1280</li> <li><strong>Augmentation:</strong> Mosaic (p=0.8), horizontal flip (p=0.3), random 90 degree rotation (p=0.05), color jitter, blur, noise</li> </ul> <h2>Evaluation</h2> <h3>Testing Data</h3> <p>LADaS test split.</p> <h3>Metrics</h3> <p>| | mAP@50 | mAP@50:95 | Precision | Recall | F1 | |---|---|---|---|---|---| | <strong>Overall</strong> | <strong>0.3864</strong> | <strong>0.2782</strong> | <strong>0.3864</strong> | <strong>0.7668</strong> | <strong>0.5139</strong> |</p> <h4>Per-Class Results</h4> <p>| Class | mAP@50:95 | Precision | Recall | F1 | |---|---|---|---|---| | AdvertisementZone | 0.0523 | 0.0679 | 0.5514 | 0.1209 | | DigitizationArtefactZone | 0.3694 | 0.5207 | 0.6964 | 0.5959 | | DropCapitalZone | 0.5432 | 0.8100 | 0.7055 | 0.7541 | | FigureZone | 0.0271 | 0.0307 | 0.3600 | 0.0567 | | FigureZone-FigDesc | 0.0000 | 0.0000 | 0.0000 | 0.0000 | | FigureZone-Head | 0.0052 | 0.0058 | 0.1800 | 0.0112 | | FormZone | 0.0094 | 0.0148 | 0.6654 | 0.0290 | | GraphicZone | 0.4883 | 0.5916 | 0.7750 | 0.6710 | | GraphicZone-Decoration | 0.2527 | 0.3592 | 0.6352 | 0.4589 | | GraphicZone-FigDesc | 0.1031 | 0.1580 | 0.5392 | 0.2444 | | GraphicZone-Head | 0.3476 | 0.6019 | 0.6275 | 0.6144 | | GraphicZone-Part | 0.1126 | 0.1424 | 0.4152 | 0.2121 | | GraphicZone-TextualContent | 0.0289 | 0.0476 | 0.3185 | 0.0828 | | MainZone-Continued | 0.4327 | 0.4796 | 0.6976 | 0.5684 | | MainZone-Date | 0.2639 | 0.4096 | 0.6220 | 0.4940 | | MainZone-Entry | 0.5473 | 0.6969 | 0.7535 | 0.7241 | | MainZone-Head | 0.4114 | 0.6113 | 0.6923 | 0.6493 | | MainZone-Lg | 0.5114 | 0.6570 | 0.7532 | 0.7018 | | MainZone-ListItem | 0.1449 | 0.1970 | 0.5843 | 0.2947 | | MainZone-Maths | 0.0196 | 0.0230 | 0.4993 | 0.0439 | | MainZone-Other | 0.4441 | 0.4862 | 0.6153 | 0.5432 | | MainZone-P | 0.6511 | 0.7357 | 0.8557 | 0.7912 | | MainZone-Signature | 0.6859 | 0.7513 | 0.9182 | 0.8264 | | MainZone-Sp | 0.6308 | 0.7466 | 0.8630 | 0.8006 | | MarginTextZone-ContinuedNotes | 0.0311 | 0.0418 | 0.5000 | 0.0772 | | MarginTextZone-ManuscriptAddendum | 0.2791 | 0.5656 | 0.5183 | 0.5409 | | MarginTextZone-Notes | 0.3839 | 0.5697 | 0.6682 | 0.6150 | | MusicZone | 0.4326 | 0.6603 | 0.6642 | 0.6623 | | NumberingZone | 0.3413 | 0.7173 | 0.5197 | 0.6028 | | QuireMarksZone | 0.2981 | 0.5588 | 0.6423 | 0.5976 | | RunningTitleZone | 0.4324 | 0.7050 | 0.6495 | 0.6761 | | StampZone | 0.4257 | 0.5657 | 0.7160 | 0.6321 | | StampZone-Sticker | 0.1798 | 0.2412 | 0.6303 | 0.3489 | | TableZone | 0.1894 | 0.2559 | 0.4828 | 0.3345 | | TableZone-Head | 0.0302 | 0.0499 | 0.4256 | 0.0893 | | TitlePageZone | 0.1386 | 0.1672 | 0.7645 | 0.2743 | | TitlePageZone-Index | 0.0486 | 0.0541 | 0.5455 | 0.0984 |</p> <h2>Citation</h2> <p><strong>BibTeX:</strong></p> <p><code>bibtex @misc{clrice2024ladaslargemultitaskdiachronic, title={LADaS -- a Large multi-task and diachronic dataset for Layout Analysis of diverse historical documents}, author={Thibault Clérice}, year={2024}, eprint={2411.10068}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2411.10068}, }</code></p>