YOLOv8 Object Detection Model for Die Bombe
Model Description
This repository contains a fine-tuned YOLOv8m object detection model trained on digitized pages of the historical satirical periodical Die Bombe.
The model is based on the Ultralytics yolov8m.pt checkpoint.
Classes
The annotation schema contains seven categories:
AdvertisementComicEditorial CartoonHeadlineIllustrationMapPhotograph
The training data contain the following numbers of annotated objects:
| Class | Training instances |
|---|---|
| Advertisement | 2,596 |
| Headline | 1,471 |
| Illustration | 465 |
| Editorial Cartoon | 140 |
| Comic | 61 |
| Photograph | 1 |
| Map | 0 |
Photograph and Map are therefore part of the annotation schema but are not meaningfully represented in the trained model.
Training Data
The model was trained on manually annotated pages of Die Bombe.
The dataset was divided into:
- 508 training pages
- 106 validation pages
The training and validation sets contain separate periodical issues.
The annotated training dataset is published separately on Zenodo:
Dataset: [Training-Dataset_Bombe_1871-1925]
DOI: [10.5281/zenodo.18757852]
Training Configuration
Training was performed with Ultralytics YOLOv8 using the following configuration:
| Parameter | Value |
|---|---|
| Base model | yolov8m.pt |
| Maximum epochs | 100 |
| Image size | 1280 |
| Batch size | 4 |
| Workers | 2 |
| Patience | 20 |
| Optimizer | SGD |
Initial learning rate (lr0) | 0.01 |
| Cosine learning-rate schedule | True |
| Mosaic augmentation | 1.0 |
| Cache | False |
| Seed | 0 |
| Deterministic | True |
Early stopping was enabled with a patience value of 20. No further improvement in mAP@0.50–0.95 was observed after epoch 17, and training stopped after epoch 37.
The best-performing checkpoint is provided as best.pt.
Evaluation
The final model was evaluated on the held-out validation set of 106 pages containing 1,004 annotated objects.
Overall Results
| Metric | Score |
|---|---|
| Precision | 0.836 |
| Recall | 0.835 |
| mAP@0.50 | 0.913 |
| mAP@0.50–0.95 | 0.722 |
Class-Specific Results
| Class | Instances | Precision | Recall | mAP@0.50 | mAP@0.50–0.95 |
|---|---|---|---|---|---|
| Advertisement | 505 | 0.930 | 0.935 | 0.967 | 0.879 |
| Comic | 10 | 0.746 | 0.900 | 0.945 | 0.567 |
| Editorial Cartoon | 33 | 0.737 | 0.818 | 0.859 | 0.788 |
| Headline | 341 | 0.953 | 0.827 | 0.961 | 0.703 |
| Illustration | 115 | 0.814 | 0.696 | 0.832 | 0.673 |
Photograph and Map do not have meaningful evaluation results because they are not sufficiently represented in the training data.
Reproducibility
The code and notebooks used for model training, evaluation, and subsequent processing are available in the Building Character GitHub repository:
Code: https://github.com/lisagollner/Building-Character_Code
The Corpus created with the workflow using this model is distributed over Zenodo:
Corpus-Title: Building Character Corpus
The model was trained using:
- Ultralytics: YOLOv8.2.0
- Python: 3.12.13
- PyTorch: 2.4.1+cu121
- GPU: NVIDIA Tesla T4
Additional training configuration is included with the model files.
License
This model is released under the GNU Affero General Public License v3.0 (AGPL-3.0).
The model was trained using the Ultralytics YOLOv8 framework. Users should consult the applicable Ultralytics and AGPL-3.0 licensing terms when reusing or redistributing the model.