Archival Faces: Detection of Faces in Digitized Historical Documents

Archival Faces Dataset This is a YOLO-formatted detection dataset for archival documents, specifically designed for facial detection and suitable for evaluation datasets or cross-validation experiments. Dataset Overview The Archival Faces dataset contains historical face images formatted for use with YOLO (You Only Look Once) object detection frameworks. Format All annotations follow the YOLO format: Each image has a corresponding .txt file with the same name Each line in the annotation file represents one face Format: class_id x_center y_center width height right_eye_x right_eye_y right_eye_visibility left_eye_x left_eye_y left_eye_visibility nose_x nose_y nose_visibility right_mouth_corner_x right_mouth_corner_y right_mouth_corner_visibility left_mouth_corner_x left_mouth_corner_y left_mouth_corner_visibility All values are coordinated normalized between 0 and 1, and visibility is always 1 class_id It is always zero since there is only one class (face) Dataset Structure The dataset is organized into 10 splits for cross-validation (k=10, k=2, k=5) and contains the following structure: ArchivalFaces/ ├── fold_0/ │ ├── images/ │ └── labels/ ├── fold_1/ │ ├── images/ │ └── labels/ ... ├── fold_9/ │ ├── images/ │ └── labels/ ... ├── fold_0.yaml ├── fold_1.yaml ... ├── k=2_fold_0.yaml ├── k=2_fold_1.yaml ... ├── k=5_fold_0.yaml ├── k=5_fold_1.yaml ... └── README.md Usage for Ultralytics YOLO model training (v8.3.92, Python 3.10) Extract ArchivalFaces.zip Change directory to ArchivalFaces/ the one that contains the dataset and requirements.txt Create a virtual environment and install all the dependencies. python -m venv .venv source .venv/bin/activate pip install -r requirements.txt Set the Ultralytics dataset to the root directory of this repository. Either change the config via the command line: yolo settings datasets_dir='/path/to/current/directory' Or add from ultralytics import settings; settings.update({"datasets_dir": "/path/to/current/directory"}) to your training script Train your model on k-fold cross-validation python training_example.py Pre-trained models All cross-validation-trained models are added under ArchivalFaces__models.zip archive. The structure of the archive follows the format: ArchivalFaces__models/ ├── {DATASET}[_{PROCESSING_DATE}_k={CROSS_VALIDATION_CONFIGURATION}]/ │ ├── {MODEL}[_fold=[FOLD]].pt/ │ └── ... └── ... DATASET can be either WiderFace or ArchivalFaces. For Archival Faces, there is also PROCESSING_DATE, either 2024-08-07 or 2024-04-04; and CROSS_VALIDATION_CONFIGURATION Which represents the number of cross-validation splits used for a specific experiment, either k=10, k=5, or k=2. Under each experiment directory, a list of pre-trained models is provided. We consistently report the model type, MODEL which can be either YOLOv8 or YOLO11, in variants N, S, M, and L. Models trained on Archival Faces also have the FOLD number that reports the fold used for model evaluation. All models can be used with the Ultralytics framework, and can be loaded with a few lines of code: from ultralytics import YOLO model = YOLO('/path/to/model.pt') Citation If you use this dataset in your research, please cite: @article{DBLP:journals/corr/abs-2504-00558, author = {Marek Vasko and Adam Herout and Michal Hradis}, title = {Archival Faces: Detection of Faces in Digitized Historical Documents}, journal = {CoRR}, volume = {abs/2504.00558}, year = {2025}, url = {https://doi.org/10.48550/arXiv.2504.00558}, doi = {10.48550/ARXIV.2504.00558}, eprinttype = {arXiv}, eprint = {2504.00558}, biburl = {https://dblp.org/rec/journals/corr/abs-2504-00558.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} } @dataset{vasko_2025_15077975, author = {Vaško, Marek and Herout, Adam and Hradiš, Michal and Dvořáková, Martina and Petr, Žabička}, title = {Archival Faces: Detection of Faces in Digitized Historical Documents }, month = mar, year = 2025, publisher = {Zenodo}, version = {1.0}, doi = {10.5281/zenodo.15077975}, url = {https://doi.org/10.5281/zenodo.15077975},} License The dataset is made available for non-commercial use (Attribution-NonCommercial 4.0 International). For commercial use, please get in touch with the authors to obtain a commercial license.