LADaS2 | D-FINE Kraken segmentation model based on COCO
Year: 2026
Type: dataset
Abstract: Overview LADaS, created by the ALMANaCH team-project at Inria, continued in partnership with other researchers, is a multidocuments diachronic layout analysis dataset. The dataset comprises 8,287 manually annotated pages spanning 1600--2024, covering both digitised and born-digital documents across diverse genres including magazines, academic papers, monographs, plays, and administrative records. The script in document is mostly Latin script, and language is mostly French with some representation of the main western academic languages. This model was trained on a single NVIDIA A100 GPU with input resolution 960*960, batch size 8 and 50 epochs. Gradient accumulation over 4 steps was used, yielding an effective batch size of 32. This variant used a random initialisation of the detection head and a COCO-pretrained backbone. Training data More information on Github. License This model is free to use for non-commercial purposes under CC-BY-NC-SA 4.0. Support This project has been orginally established and funded as part of the DEFI COLaF (2023-2027) with help of the European project ATRIUM and FNS-granted project The Geographic Horizon of writers. Publications Juliette Janes, Sarah Bénière, Benjamin Kiessling, Lucence Ing, Éric Astier, et al.. LADaS: A Diachronic Multi-Type Dataset for Document Layout Analysis from the 17th Century to the Present. 8th International Workshop on Historical Document Imaging and Processing (HIP’26), The 20th International Conference on Document Analysis and Recognition (ICDAR), Sep 2026, Vienne, Austria. ⟨hal-05638101⟩
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223661879