Document Layout Analysis

PubLayNet is a dataset for document layout analysis by automatically matching the XML representations and the content of over 1 million PDF articles that are publicly available on PubMed Central. The size of the dataset is comparable to established computer vision datasets, containing over 360 thousand document images, where typical document layout elements are annotated. Source: PubLayNet: largest dataset ever for document layout analysis

The RVL-CDIP dataset consists of scanned document images belonging to 16 classes such as letter, form, email, resume, memo, etc. The dataset has 320,000 training, 40,000 validation and 40,000 test images. The images are characterized by low quality, noise, and low resolution, typically 100 dpi. Source: Towards a Multi-modal, Multi-task Learning based Pre-training Framework for Document Representation Learning Image Source: https://www.cs.cmu.edu/~aharley/rvl-cdip/

A benchmark dataset that contains 500K document pages with fine-grained token-level annotations for document layout analysis. DocBank is constructed using a simple yet effective way with weak supervision from the \LaTeX{} documents available on the arXiv.com. Source: DocBank: A Benchmark Dataset for Document Layout Analysis

Various documents dataset. Each of the 65 documents includes scanned ground truth images, both hard and easy distorted photos, and document-centered cropped images. Image Source: DocUNet: Document Image Unwarping via a Stacked U-Net

The database consists of 150 annotated pages of three different medieval manuscripts with challenging layouts. Furthermore, we provide a layout analysis ground-truth which has been iterated on, reviewed, and refined by an expert in medieval studies.

The DSSE-200 is a complex document layout dataset including various dataset styles. The dataset contains 200 images from pictures, PPT, brochure documents, old newspapers and scanned documents.

HJDataset is a large dataset of Historical Japanese Documents with Complex Layouts. It contains over 250,000 layout element annotations of seven types. In addition to bounding boxes and masks of the content regions, it also includes the hierarchical structures and reading orders for layout elements. The dataset is constructed using a combination of human and machine efforts. Source: A Large Dataset of Historical Japanese Documents with Complex Layouts Image Source: https://dell-research-harvard.github.io/HJDataset/

We present the VIS30K dataset, a collection of 29,689 images that represents 30 years of figures and tables from each track of the IEEE Visualization conference series (Vis, SciVis, InfoVis, VAST). VIS30K’s comprehensive coverage of the scientific literature in visualization not only reflects the progress of the field but also enables researchers to study the evolution of the state-of-the-art and to find relevant work based on graphical content. We describe the dataset and our semi-automatic collection process, which couples convolutional neural networks (CNN) with curation. Extracting figures and tables semi-automatically allows us to verify that no images are overlooked or extracted erroneously. To improve quality further, we engaged in a peer-search process for high-quality figures from early IEEE Visualization papers.

The D4LA dataset is a diverse benchmark for document layout analysis (DLA) derived from the RVL-CDIP dataset. It focuses on 12 document types with rich layouts, each represented by approximately 1,000 manually annotated images, while filtering out noisy, handwritten, artistic, or text-scarce images. The dataset defines 27 detailed layout categories, including DocTitle, ListText, Header, Table, Equation, and Footer, among others, catering to real-world applications.

U-DIADS-Bib is a proprietary dataset developed through the collaboration of computer scientists and humanities at the University of Udine. It is composed of 200 images, 50 for each of the 4 different manuscripts that characterize it. These handwritten books were selected in collaboration with humanist partners considering both the complexity of their layout and the presence of significant and semantically distinguishable elements. In particular, the images of the four manuscripts were collected from the digital library Gallica. All manuscripts are Latin and Syriac Bibles published between the 6th and 12th centuries A.D.

• Revision: v1.0.0-full-20210527a - DOI: 10.5281/zenodo.4817662 - Authors: J. Chazalon, E. Carlinet, Y. Chen, J. Perret, C. Mallet, B. Duménieu and T. Géraud - Official competition website: https://icdar21-mapseg.github.io/ This is the dataset of the ICDAR 2021 Competition on Historical Map Segmentation (“MapSeg”). This competition ran from November 2020 to April 2021. ## Motivation This competition aims as encouraging research in the digitization of historical maps. In order to be usable in historical studies, information contained in such images need to be extracted. The general pipeline involves multiples stages; we list some essential ones here: - segment map content: locate the area of the image which contains map content; - extract map object from different layers: detect objects like roads, buildings, building blocks, rivers, etc. to create geometric data; - georeference the map: by detecting objects at known geographic coordinate, compute the transformation to turn geometric objects into geographic ones (which can be overlaid on current maps). ## Tasks The tasks we propose simulate the three essential digitization steps we just mentioned. ### Task 1: “Detect Building Blocks” This task is the flagship of this competition. Given a fragment of map sheet image focused on map content, you need to detect the building blocks. Building blocks are symbolized by a thick line. They do not overlap between themselves, but many other elements can perturb their detection: - special buildings (hatched areas) can be included in building blocks (sometimes they cover the building block completely); - text can be overlaid on lines; - those maps contain many lines which need to be filtered out (internal building structures, railways, rivers, gardens…). Expected output for this task is a binary mask indicating for each pixel whether it belongs to a building block or not. Evaluation tools also tolerate a label map in TIFF format, where each pixel is labelled with the identifier of the shape it belongs to, using an INT16. To extract shapes from the binary mask, 4-connectivity is used (hence the background has 8-connectivity). ### Task 2: “Segment Map Area” This tasks is the equivalent of text area detection for OCR: given the image of a complete map sheet, you need to segment the area which contains map content. This area is usually well separated from the other elements (title, legend, scale…) by several frames but sometimes map contents exceeds the frame for some large objects. While most of the area is delineated by straight lines, some objects were drawn outside the frame on several sheets. We decided to segment each of those regions as closely as possible. Expected output for this task is a binary mask indicating for each pixel whether it belongs to the map area or not. ### Task 3: “Locate Graticule Lines Intersections” This task is essential to the georeferencing of the map: graticule lines are lines which indicate the North/South/East/West coordinates relative to the reference point. Their intersection points are very useful to provide key points for the registration of the map image. Given the image of a complete map sheet, you need to locate the intersection points of such lines. These lines usually cover the map content from left to right or from top to bottom but beware: - due to document aging paper sheets are not flat anymore and lines are not straight; - lines may be in diagonal for some areas; - lines can be overlaid with many other objects. Expected output for this task is a list of coordinates (in image referential, i.e. 0,0 at top left, x-axis pointing to the right and y-axis pointing downward).

We compiled a new dataset (the PERO layout dataset) that contains 683 images from various sources and historical periods with complete manual text block, text line polygon and baseline annotations. The included documents range from handwritten letters to historic printed books and newspapers and contain various languages including Arabic and Russian. Part of the PERO dataset was collected from existing datasets and extended with additional layout annotations (cBAD, IMPACT and BADAM). The dataset is split into 456 training and 227 testing images.

The UrduDoc Dataset is a benchmark dataset for Urdu text line detection in scanned documents. It is created as a byproduct of the UTRSet-Real dataset generation process. Comprising 478 diverse images collected from various sources such as books, documents, manuscripts, and newspapers, it offers a valuable resource for research in Urdu document analysis. It includes 358 pages for training and 120 pages for validation, featuring a wide range of styles, scales, and lighting conditions. It serves as a benchmark for evaluating printed Urdu text detection models, and the benchmark results of state-of-the-art models are provided. The Contour-Net model demonstrates the best performance in terms of h-mean. The UrduDoc dataset is the first of its kind for printed Urdu text line detection and will advance research in the field. It will be made publicly available for non-commercial, academic, and research purposes upon request and execution of a no-cost license agreement. To request the dataset and for more information and details about the UrduDoc , UTRSet-Real & UTRSet-Synth datasets, please refer to the Project Website of our paper "UTRNet: High-Resolution Urdu Text Recognition In Printed Documents"

opendateset