README.md
Matthieu Pélingre, Salvatore Tabbone · Recherche Data Gouv France · 2026
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Recherche Data Gouv France · OpenAlex · dataset · 2026
Abstract
This deposit contains Historical Postcards Dataset (COCO) — v2.0 (2026), a Common Objects in COntext (COCO) format dataset of historical postcard images and structured annotations intended for text and postal markings detection. Compared to v1, all text regions in the Train and Test sets now carry manually verified transcriptions — for printed text (printed_text), handwritten text (handwritten_text) and scene text (scene_text) — together with text orientation, language, and an illegibility flag (marking regions that could not be read).
The synthetic set (Synth) provides automatic detections using fine-tuned YOLOv12x, as well as transcriptions for the three text types, produced with HunyuanOCR. This makes the dataset suitable for detection and historical OCR benchmarking. Transcriptions of postal markings will be added in future versions. The COCO format allows the use of the COCO API (or pycocotools for Python). Contents & purpose 4,293 postcard images (Image’Est archive, Grand-Est region, France; period ca. 1899–1930) with COCO annotations for text and postal markings detection.
The dataset is intended for training/evaluating text detection, OCR pipelines, and postal markings recognition for cultural-heritage research. The first version of the dataset was presented at the 7th workshop on analySis, Understanding and proMotion of heritAge Contents (7th SUMAC @ ACM Multimedia 2025), on 27 October 2025 in Dublin, Ireland; the related paper describes the dataset and the detection pipeline.
The v2 companion paper will be presented at the 20th International Conference on Document Analysis and Recognition (20th ICDAR 2026), on 2 September 2026 in Vienna, Austria; it benchmarks current VLMs against our specialized pipeline for both the detection and transcription of the text regions. What's new in v2.0 Manual transcriptions extended to every text region of the Train and Test sets (printed, handwritten and scene text), not only printed text. Synthetic transcriptions (Synth set) regenerated with HunyuanOCR (previously EasyOCR), for the three text types. New language attribute (fr, de, en) and new illegible attribute on text annotations. Removed the confidence attribute.
Minor annotation corrections. See the README.md file for the full changelog and annotation schema.
Provided archives Historical_Postcards_Dataset_v2-Train2026.zip — manually annotated set (train split of the detection experiments, used in 5-fold cross-validation) annotations-Historical_Postcards_Dataset_v2-Train2026.zip — annotations only Historical_Postcards_Dataset_v2-Test2026.zip — manually annotated set (held-out test split of the detection experiments) annotations-Historical_Postcards_Dataset_v2-Test2026.zip — annotations only Historical_Postcards_Dataset_v2-Synth2026.zip — set with synthetic (automatic) annotations annotations-Historical_Postcards_Dataset_v2-Synth2026.zip — annotations only The Train/Test split relates to the detection experiments; the transcription evaluation in the v2 companion paper was carried out on a manually validated subset of text regions spanning both splits (see the README.md file).
This subdivision also facilitates importation into CVAT. More information in the README.md file. File structure & Annotation schema See the README.md file for more details. Acknowledgments We would like to thank Image'Est for making historical postcards data available, the Grand Est Region, France who supported this work, and our annotators.
Open the resolved locator · 10.57745/rl9oit
Metadata licence: OpenAlex metadata (CC0)