Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,70 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- object-detection
|
| 5 |
+
- instance-segmentation
|
| 6 |
+
- layout-analysis
|
| 7 |
+
- document-analysis
|
| 8 |
+
- manuscripts
|
| 9 |
+
- segmonto
|
| 10 |
+
- leibniz
|
| 11 |
+
- historical-documents
|
| 12 |
+
- rf-detr
|
| 13 |
+
- dinov2
|
| 14 |
+
pipeline_tag: image-segmentation
|
| 15 |
+
library_name: rfdetr
|
| 16 |
+
metrics:
|
| 17 |
+
- mAP
|
| 18 |
---
|
| 19 |
+
|
| 20 |
+
# Zone detection model for (17th–18th c.) manuscripts, trained on Leibniz's manuscripts using SegmOnto ontology - V2.
|
| 21 |
+
|
| 22 |
+
This model was developed by Denisa-Florina BUMBA as part of the ERC Philiumm project (Grant ADG No. 101020985), at the Laboratoire SPHERE, Université Paris Cité – CNRS, France.
|
| 23 |
+
|
| 24 |
+
Fine-tuned RF-DETR-Seg (Nano) for instance segmentation and zone detection on scanned pages of Leibniz's manuscripts. The model detects zones such as main text, margin text, numbering, diagrams, digitization artifacts, and several classes of mathematical formulae using the SegmOnto controlled vocabulary, in order to support automatic transcription in the digital humanities. It is trained specifically on Leibniz's manuscripts, but may be applied to other historical manuscripts with similar layout.
|
| 25 |
+
|
| 26 |
+
**Previous version:** [v1](https://huggingface.co/DenisaBumba/rfdetr-instance-segmentation-leibniz-manuscripts)
|
| 27 |
+
|
| 28 |
+
What's new in v2 :
|
| 29 |
+
|
| 30 |
+
- corrected annotations, particularly for mathematical formulae zones.
|
| 31 |
+
- increased the training resolution from 720 px to 1272 px.
|
| 32 |
+
|
| 33 |
+
## Classes
|
| 34 |
+
|
| 35 |
+
| Class | Description |
|
| 36 |
+
|---|---|
|
| 37 |
+
| `MainZone` | Main body text region |
|
| 38 |
+
| `MarginTextZone` | Marginal additions |
|
| 39 |
+
| `NumberingZone` | Page numbering |
|
| 40 |
+
| `DigitizationArtefactZone` | Digitization artefacts |
|
| 41 |
+
| `GraphicZone-figure` | Diagrams, other graphic zones |
|
| 42 |
+
| `GraphicZone-formula` | Covers all mathematical formulae |
|
| 43 |
+
| `MathsRegion-complex` | Complex/multi-line formulae |
|
| 44 |
+
| `MathsRegion-standard` | Inline formulae within text |
|
| 45 |
+
| `MathsRegion-strikethrough` | Struck-through/deleted formulae |
|
| 46 |
+
|
| 47 |
+
## Training Data
|
| 48 |
+
|
| 49 |
+
| Split | Images | Percentage |
|
| 50 |
+
|---|---|---|
|
| 51 |
+
| Train | 1,002 | 94% |
|
| 52 |
+
| Validation | 40 | 4% |
|
| 53 |
+
| Test | 18 | 2% |
|
| 54 |
+
| **Total** | **1,060** | 100% |
|
| 55 |
+
|
| 56 |
+
## Performance
|
| 57 |
+
|
| 58 |
+
Results on the validation set.
|
| 59 |
+
|
| 60 |
+
| Checkpoint | Epoch | segm mAP@50:95 |
|
| 61 |
+
|---|---|---|
|
| 62 |
+
| `checkpoint_best_regular.pth` | 40/60 | 0.4670 |
|
| 63 |
+
| `checkpoint_best_ema.pth` | 40/60 | **0.4700** |
|
| 64 |
+
|
| 65 |
+
## eScriptorium pipeline
|
| 66 |
+
|
| 67 |
+
Use the [RF-DETR-Seg to XML Page pipeline](https://gitlab.com/eman8/scripts/scripts-pour-le-pretraitement-des-corpus-sur-EMAN/scripts-pour-le-pretraitement-des-corpus-pour-escriptorium)
|
| 68 |
+
|
| 69 |
+
## Funding
|
| 70 |
+
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme, Advanced Grant ADG No. 101020985.
|