Datasets per Publication

A new database with handwritten Arabic script. It is based on five books written by different writers from the years 1088-1451. We took 668 pages from these five books, and fully annotated them on the sub-word level. For each page we manually applied bounding boxes on the different sub-words and annotated the sequence of characters. It consists of 159,149 sub-word appearances consisted of 326,289 characters out of a vocabulary of 5,509 forms of sub-words. The database is described in detail and is designed for training and testing recognition systems for handwritten Arabic sub-words. This database is available for the purpose of research, and we encourage researchers to develop and test new methods using our database.

For tutorials to understand the annotation format, extraction, and segmentation of this dataset, please refer here. In order to get more information regarding the dataset click here, and to read about the basic tracks in order to conduct research on then click here.

The dataset contains 419 handwritten flowcharts using the inkML standard format.
35 writers have contributed. The train set (248 samples) contains 14 different patterns and the test set (171 samples) uses 14 other flowchart patterns.
 

The collection contains simultaneously acquired online and offline samples. The online dataset is called "NFI-online" and was processed and segmented by Louis Vuurpijl. The offline dataset is called "NFI-offline" and was scanned by Vivian Blankers from the NFI. The offline signatures were automatically segmented and visually inspected by Louis Vuurpijl. Processing of the segmented signatures was performed by Katrin Franke. The collection contains:

The NLDCC-online signature collection contains in total 1953 online and 1953 offline signature files.

This is the dataset of the ICDAR 2013 - Gender Identification from Handwriting competition. If you use this database, please consider citing it as in [1].

This dataset is a subset of the QUWI dataset [2]. In sum, a total of 475 writers produced 4 handwritten documents: the first page contains an Arabic handwritten text which varies from one writer to another, the second page contains an Arabic handwritten text which is the same for all the writers, the third page contains an English handwritten text which varies from one writer to another and the fourth page contains an English handwritten text which is the same for all the writers.

Images have been acquired using an EPSON GT-S80 scanner, with a 300 DPI resolution. Images were provided in
JPG uncompressed format. The training set consists of the first 282 writers for which the genders are provided.
Participants were asked to predict the gender of the remaining 193 writers.

In addition to images, features extracted from the data were also provided in order to make the competition accessible by people without image processing skills. Those features are described in [3].

[1] Hassaïne, Abdelâali, et al. "ICDAR 2013 Competition on Gender Prediction from Handwriting." Document Analysis and Recognition (ICDAR), 2013 12th International Conference on. IEEE, 2013.

[2] Hassaïne, Abdelaali, et al. "The ICDAR2011 Arabic writer identification contest." Document Analysis and Recognition (ICDAR), 2011 International Conference on. IEEE, 2011.

[3] Al Maadeed, Somaya, and Abdelaali Hassaine. "Automatic prediction of age, gender, and nationality in offline handwriting." EURASIP Journal on Image and Video Processing 2014.1 (2014): 1-10.

SOURCE WEBSITES:

Statistics Canada  http://www.statcan.gc.ca/start-debut-eng.html

The World Bank,   http://www.worldbank.org/

Statistics Norway,    https://www.google.com/?gws_rd=ssl#q=Statistics+Norway

Statistics Finland, http://www.stat.fi/index_en.html

US Department of Justice,    https://www.justice.gov/

Geohive,   http://www.geohive.com/

US Energy Information Administration,    https://www.eia.gov/

US Census Bureau. http://www.census.gov/

 DATA COLLECTION:  About 1000 tables were collected from international statistical websited by DocLab graduate students in2009-2010. A perceptually random subset of 200 of these tables was converted (mostly from HTML) via EXCEL into CSV files and stored at DocLab. The original HTML files were not retained and their URLs were kept somewhat haphazardly. Many of these tables can still be found by a web search, but others have been modified, corrected, or updated on the original sites. The dataset now consists only of the 200 CSV file representation of 200 web tables.

GROUND TRUTH: The four critical cells (top-left and bottom right of the stub-header and top-left and bottom-right of the data region) were entered with an interactive tool (VeriClick) in 2011. Several GT errors were subsequently found during segmentation experiments and corrected. Researchers may still disagree on 2-3 tables on whether a particular row or column (unusual contents or units) belong to the data region.  The GT for all 200 tables is a single CSV file that allows adding to it and verifying the results of a header segmentation experiment.

AcTiV is the first publicly accessible annotated dataset designed to assess the performance of different Arabic VIDEO-OCR systems. The database has been named AcTiV for Arabic Text in Video. The challenges that are addressed by AcTiV-database are in text patterns variability and presence of complex background with various objects resembling text characters. AcTiV enables users to test their systems’ abilities to locate, track and read text objects in videos. The actual version of the dataset includes 80 videos collected from 4 different Arabic news channels. In the present work, two types of video stream were chosen: Standard-Definition (720x576, 25 fps) and High-Definition (1920x1080, 25fps).We mainly focus on text displayed as overlay in news video, which can be classified into two types: static text and dynamic one.

Two sub-datasets are created from the AcTiV database: Activ-D (D for detection) and  Activ-R (R for recognition). AcTiV-D represents a sub-dataset of nonredundant frames used to measure the performance of single-frame based methods to detect/localize text regions in still HD/SD images. AcTiV-R is a sub-dataset of cropped images used to measure the performance of Arabic OCR systems to read texts in video frames.

Please find hereafter the download links of the detection and recognition datasets.

The Bentham collection consists of a set of images of a collection of works on law and moral philosophy written by the philosopher Jeremy Bentham.

The selected subset has been written by several hands (Bentham himself and his secretaries) and entails significant varibilities and difficulties regarding the quality of text images and writting styles. Training and test data were provided in the form of carefully segmented line images, along with the corresponding transcripts.

This dataset is free available for research purposes and it is provided into two parts: the images and the GT. The GT includes information about the  layout and the transcription at line level of each image in PAGE format. 

Keywords

Forensic analysis, signatures, handwriting, off-line, on-line, verification, identification, evaluation, likelihood ratios

Description

Spectral analysis of writing materials is of great importance for the study and analysis of ancient documents. Multi-spectral (MS) imaging represents an innovative and non-destructive technique for the analysis of materials. For this purpose, we collected a Multispectral image database of ancient handwritten letters.

The SMADI database consists of 240 multispectral images of 30 real historical handwritten letters. All ancient manuscripts were written using iron gall ink and date from the 17th to the 20th century. Original documents were borrowed from Quebec’s national library (BAnQ Bibliotheque et Archives nationales du Quebec) and have been imaged using a CROMA CX MSI camera, producing 8 images for each document. A total of 240 images of real documents were captured, calibrated (illumination) and registrated.

This site provides an assessment of the quality-time performance of binarization algorithms.

Over 5.5M synthetic, 708 photographed  and 116  "real-world" documents are used in such assessment. The results of binarization are shown in statistical tables together with time elapsed. The binary images may be obtained on-demand. The current version of the platforms presents the results of 20 "classical" global binarization algorithms. The platform is being extended to include more recent algorithms.

The Total-Text dataset consists of 1555 images collected from 1) internet sources, 2) real scenery (camera or phone captured) from many countries like Malaysia, Vietnam, Japan, Canada, etc. It was annotated for three different tasks: detection, recognition, and segmentation. English is the only annotated language, the rest of the language was labeled as 'do not care' region.

This dataset contains the test and training set for the ICDAR2017 Competition on Historical Document Writer Identification (Historical-WI).

The dataset used in this competition consists of 3600 handwritten pages originating from 13th to 20th century. It contains manuscripts from 720 different writers where each writer contributed five pages. In addition, a training set is also released which contains of 394 independent writers which 3 pages each.

Example and evaluation dataset used for the ICDAR2017 Competition on Recognition of Documents with Complex Layouts, as it was made available to the participants of the competition.

Contains colour images and page layout ground truth of magazine pages and technical articles.

Example and evaluation dataset used for the ICDAR2017 Competition on Recognition of Early Indian printed Documents, as it was made available to the participants of the competition.

Challenge to find an optimal solution for accurately and automatically transcribing the Bengali books and Quarterly Lists, to form a unique dataset that can be used with computational tools and methods, and to enable full-text search and discovery.

The dataset is multilingual (50% French, 50% English, with maybe few expressions borrowed from other languages such as Latin). It consists in texts published in monographs and periodicals during the last four centuries. It accounts more than 12 million characters, and includes both noisy OCR-ed texts and the corresponding Gold-Standard (GS) which has been aligned at the character level.

The documents come from different collections (e.g. BnF, British Library) supported by various projects (e.g. Europeana Newspapers, IMPACT, Gutenberg, Perseus, Wikisource and Bank of wisdom).

Error rates, and thus difficulty, vary according to the nature of the documents. In both English and French parts of the dataset, historical newspapers, which have been reported to be especially challenging, represent more than 25% of the data.

Important notice, in the archive :

 Download link:
https://drive.google.com/open?id=1-pNT00vvIqh0ss_5b2aHo-nG8advaFJi

This is a dataset of chart images created from real data sources using matplotlib.  There are 4 basic types of charts: Bar, Line, Scatter, Box.  There are several tasks associated with this dataset including:

1) Chart Classification

2) Text Detection and Recognition

3) Text Role Classification

4) Axis Analysis

5) Legend Analysis

6a) Plot Element Detection and Classification

6b) Data Extraction

These tasks form a pipeline to recover the data used to generate the chart.

Later tasks take as input the ideal output of the previous tasks.

For more information, please visit https://chartinfo.github.io/

The dataset may be downloaded from https://github.com/adobe-research/CHART-Synthetic/releases

For this competition, the evaluation set consisted of 85 images. These included ten new scans taken from IEEE Spectrum magazines and 75 images selected from the PRImA Layout Analysis dataset as a representative sample ensuring a balanced presence of different issues affecting layout analysis and OCR. Such issues include the presence of non-rectangular shaped regions, varying text column widths, varying font sizes, presence of separators and regions of “reverse video” text (light-coloured text on a dark background). The presence of running headers and captions of illustrations/photographs in addition to the main body of text, pose difficulties in the identification of the correct reading order of the page.

For the most part, the scanned images contain single column lines of text, with a small amount containing illustrations as well as text. Some pages contain marginal data such as numbers, handwritten notes, and decorative frames. The evaluation set consisted of 56 images as a representative sample ensuring a balanced presence of different issues affecting layout analysis and OCR. Such issues include non-straight text lines, show-through or bleed-through, faded ink, decorations, the presence non-rectangular shaped regions, varying text column widths, varying font sizes, presence of separators and various aging- and scanning-related issues. In addition to the evaluation set, 25 representative images were selected as the example set that was provided to the authors with ground truth. 

This is a dataset of manually annotated chart images extracted from the Pub Med Central Open Access set.  There are 4 basic types of charts: Bar, Line, Scatter, Box.  There are several tasks associated with this dataset including:

1) Chart Classification

2) Text Detection and Recognition

3) Text Role Classification

4) Axis Analysis

5) Legend Analysis

A total of 4242 images have been annotated for Task 1, 200 for Task 2 and 200 for Tasks 3 to 5. For more information, please visit https://chartinfo.github.io/ or contact: kennydav@buffalo.edu

The dataset consists of 1172 Chinese document images mainly written in Chinese traditional Han script. The document images have been taken from different books.  The dataset is developed aiming historical documents to develop robust systems for historical document analysis. In this direction, there will be a competition named Historical Document Reading Challenge on Large Chinese Structured Family Records, in short ICDAR 2019 HDRC Chinese on this database. The objective behind this competition is to boost the research on historical document analysis. The focus of the competition is to recognize and analyze the layout, and finally detect and recognize the text lines and characters of the documents in this database.

Please follow the article 

@inproceedings{simistira2019icdar2019hdrc,

archivePrefix = {arXiv},

arxivId = {1903.03341},

eprint = {1903.03341},

author = {Saini, Rajkumar and Dobson, Derek and Morrey, Jon and Liwicki, Marcus and Simistira Liwicki, Foteini},

title = {{ICDAR 2019 Historical Document Reading Challenge on Large Structured Chinese Family Records}},

booktitle={{to appear in 15th International Conference on Document Analysis and Recognition (ICDAR)}},

year = {2019},

month = {mar}, }

Dataset page on Zenodo: https://zenodo.org/record/3515403

The proposed dataset accounts for 22M OCR-ed characters (754 025 tokens) along with the corresponding groundtruth, with an unequally share of 10 European languages.

The digitized documents come from different collections available, among others, in national libraries or universities. The corresponding GT comes from initiatives such as HIMANIS, IMPACT, IMPRESSO, Open data of National Library of Finland, 4HistOCR and RECEIPT.

Degraded documents sometimes result in highly noisy OCR output and thus cannot reasonably be fully aligned with their GT. The unaligned sequences have not been included in the presented statistics (e.g. number of characters and error rates).

Error rates vary according to the nature and the state of degradation of the documents. Historical books for example, due to their complex layout and their original fonts have been reported to be especially challenging for OCR engines with up to 50% of wrongly detected characters in some documents.

Pinkas dataset is created from a historical Hebrew manuscript that contains records of Jewish communities in Europe in the early modern period (c. 1500-1800). The dataset consists of 30 pages digitized by full color digital images in JPG format with high resolution. The pages  exhibit numerous degradations, complex layout and different handwritings. They are written in a mixture of Medieval Hebrew by different writers. The writers were not professional scribes. This adds additional challenge, since very often the same letters are written in different shapes.

The ground truth is available in PAGE format . Each page is labeled at word, line and page levels.

An official train and test set partition is defined and three word spotting methods are used to set the baselines [1].

This dataset contains the training and test set used in the ICDAR 2019 Competition on Image Retrieval for Historical Handwritten Documents.

This competition investigates the performance of large-scale retrieval of historical document images based on writing style. Based on large image data sets provided by cultural heritage institutions and digital libraries, providing a total of 20 000 document images representing about 10 000 writers, divided in three types: writers of (i) manuscript books, (ii) letters, (iii) charters and legal documents. We focus on the task of automatic image retrieval to simulate common scenarios of humanities research, such as writer retrieval.

The training data set encompasses images from (i) Letters A, where each writer contributed one or three images; (ii) Manuscripts, where each writer was represented by five consecutive images from a single book. In total, it contains 300 writers contributing one page, 100 writers contributing three pages, and 120 writers contributing five pages resulting in 1200 images of 520 writers.

The test data set contains 20 000 images: About 7 500 pages stem from isolated documents (partially anonymous writers, contributing one page each), and about 12 500 pages are from writers that contributed three or five pages.

For more information or in case you use this dataset, please cite:

V. Christlein, A. Nicolaou, M. Seuret, D. Stutzmann, A. Maier: "ICDAR 2019 Competition on Image Retrieval for Historical Handwritten Documents", in 15th International Conference on Document Analysis and Recognition, 2019, Sydney, Australia

This package provides training and test data from the competitions CROHME 2011, 2012, 2013, 2014, 2016 and 2019.

This dataset, named Dset_Mix, is designed to conduct experiments and evaluate the performance of methods for clustering handwritten mathematical expressions (HMEs). The idea to form this dataset is to simulate handwritten mathematical answers (HMAs) that come from a mathematical examination (only the final answer, not the whole solution). The dataset provides simultaneously online and offline patterns.

Dset_Mix has a total of 2000 HMEs in which 263 patterns are collected from the CROHME 2016 and 1,737 patterns are synthesized by utilizing handwriting symbols also from CROHME 2016. The method of synthetic data generation we used is presented in the following paper.

“Khanh Minh Phan, Vu Tran Minh Khuong, Huy Quang Ung, and Masaki Nakagawa, "Generating Synthetic Handwritten Mathematical Expressions from a LaTeX Sequence or a MathML Script," Proc. 15th International Conference on Document Analysis and Recognition, pp. 922-927, Sydney, Australia, 2019.”

The objective of the clustering HMEs is to group the same answers together.

PubLayNet is a large dataset of document images from PubMed Central Open Access Subset. Each document’s layout is annotated with both bounding boxes and polygonal segmentations.

This is the dataset of the ICDAR 2021 Competition on Historical Map Segmentation (“MapSeg”).
This competition ran from November 2020 to April 2021.
Evaluation tools are freely available but distributed separately.

Download the dataset on Zenodo:

Official competition website: https://icdar21-mapseg.github.io/

The competition report can be cited as:

Joseph Chazalon, Edwin Carlinet, Yizi Chen, Julien Perret, Bertrand Duménieu, Clément Mallet, Thierry Géraud, Vincent Nguyen, Nam Nguyen, Josef Baloun, Ladislav Lenc, and Pavel Král, "ICDAR 2021 Competition on Historical Map Segmentation", in Proceedings of the 16th International Conference on Document Analysis and Recognition (ICDAR'21), September 5-10, 2021, Lausanne, Switzerland.

BibTeX entry:

@InProceedings{chazalon.21.icdar.mapseg, author = {Joseph Chazalon and Edwin Carlinet and Yizi Chen and Julien Perret and Bertrand Duménieu and Clément Mallet and Thierry Géraud and Vincent Nguyen and Nam Nguyen and Josef Baloun and Ladislav Lenc and and Pavel Král}, title = {ICDAR 2021 Competition on Historical Map Segmentation}, booktitle = {Proceedings of the 16th International Conference on Document Analysis and Recognition (ICDAR'21)}, year = {2021}, address = {Lausanne, Switzerland}, }

We thank the City of Paris for granting us with the permission to use and reproduce the atlases used in this work.

The images of this dataset are extracted from a series of 9 atlases of the City of Paris produced between 1894 and 1937 by the Map Service (“Service du plan”) of the City of Paris, France, for the purpose of urban management and planning. For each year, a set of approximately 20 sheets forms a tiled view of the city, drawn at 1/5000 scale using trigonometric triangulation.

Sample citation of original documents:

Atlas municipal des vingt arrondissements de Paris. 1894, 1895, 1898, 1905, 1909, 1912, 1925, 1929, and 1937. Bibliothèque de l’Hôtel de Ville. City of Paris. France.

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:

Task overview

Please refer to the enclosed README.md file or to the official website for the description of tasks and file formats.

Evaluation metrics and tools

Evaluation metrics are described in the competition report and tools are available at https://github.com/icdar21-mapseg/icdar21-mapseg-eval and should also be archived using Zenodo.

The SVC2021_EvalDB database is a novel database specifically acquired for the ICDAR 2021 Competition on On-Line Signature Verification (SVC 2021) and also used in SVC On-Going Competition. Two acquisition scenarios are considered: office and mobile scenarios.

Office scenario: on-line signatures from 75 total subjects were acquired using a Wacom STU-530 device with the stylus as writing input. Regarding the acquisition protocol, the device was placed on a desktop and subjects were able to rotate it in order to feel comfortable with the writing position. It is important to highlight that the subjects considered in the acquisition of SVC2021_EvalDB are different compared to the ones considered in the DeepSignDB database. Signatures were collected in two separated sessions with a time gap between them of at least 1 week. For each subject, there are 8 total genuine signatures (4 signatures/session) and 16 skilled forgeries (8 signatures/type) performed by four different subjects in two different sessions. Regarding the skilled forgeries, both static and dynamic forgeries were considered in the first and second acquisition sessions, respectively. Information related to X and Y spatial coordinates, pressure, and timestamp is recorded for the Wacom device. In addition, pen-up trajectories are also available.

Mobile scenario: on-line signatures from 119 total subjects were acquired using the same acquisition framework considered in MobileTouchDB. Regarding the acquisition protocol, we implemented an Android App and uploaded it to the Play Store in order to study an unsupervised mobile scenario. This way all subjects could download the App and use it on their own devices without any kind of supervision, simulating a practical scenario in which subjects can generate touchscreen on-line signatures in any possible scenario, e.g., standing, sitting, walking, indoors, outdoors, etc. As a result, 94 different smartphone models from 16 different brands were collected during the acquisition. Regarding the acquisition protocol, between four and six separated sessions in different days were considered with a total time gap between the first and last session of at least 3 weeks. For each subject, there are at least 8 total genuine signatures (2 signatures/session) and 16 skilled forgeries (8 signatures/type) performed by four different subjects. Regarding the skilled forgeries, both static and dynamic forgeries were considered, similar to the office scenario. Information related to X and Y spatial coordinates, and timestamp is recorded for all devices. Pen-up information is not available in this case.

Reference

R. Tolosana, R. Vera-Rodriguez, C. Gonzalez-Garcia, et al., “ICDAR 2021 Competition on On-Line Signature Verification”, in Proc. International Conference on Document Analysis and Recognition, ICDAR, Lausanne, Switzerland, 2021.  

The Handwritten Chess Scoresheet or HCS dataset is an annotated collection of chess scoresheet images from live chess events. It was created to aid handwriting recognition research of chess moves to automate scoresheet digitization process. This dataset contains 158 games comprised of 215 pages of scoresheets digitized using a standard cellphone camera in natural lighting conditions. These images are tightly cropped, and a standard corner detection-based transformation is applied to eliminate perspective distortion. The headers and footers were also cropped out from each image in order to maintain player anonymity. Each scoresheet possesses a total of 120 move boxes, though in most cases a given scoresheet contains some blank boxes.

Also attached (please check the "Ground Truth" section) are two text files which contain ground truth data for the scoresheets. A folder named "empty scoresheets" contains a few scoresheets with no handwriting on them. This is included in the dataset intentionally to provide a format for the chess scoresheets used, and to help with extracting individual text boxes from scoresheet images.

Authors:

Licence agreement: authors are asked to download, fill and sent back to the authors this Licence Agreement Form.

If you use this dataset, please consider citing the following paper:

LOPES JUNIOR, C. A. M., NEVES JUNIOR, R. B., BEZERRA, B. L. D., TOSELLI, A. H., IMPEDOVO, D.: Competition on components segmentation task of document photos. In: International Conference on Document Analysis and Recognition (ICDAR), pp. 1–15. Springer Nature, (2021). https://doi.org/10.1007/978-3-030-86337-1_45.

The dataset is composed of ID documents images from the following Brazilian document types (acronym in Portuguese):

National Driver's License (CNH),
Natural Persons Register (CPF), and
General Registration (RG).

The documents images were captured through different cell phone cameras at different resolutions. Documents that appear in the images are characterized by having different textures, colors, and lighting, on different real-world backgrounds with non-uniform patterns. 

As the personal information contained in the documents cannot be disclosed, the original data has been replaced by synthetic data. We generate false information for all data present in the document (such as name, date of birth, affiliation, numerical data, among others). On the other hand, handwritten signatures were acquired from the MCYT and GPDS datasets, which were synthetically incorporated into the images of the identity documents.

If you use this dataset, please consider citing the following paper:

SOARES, A. G. ; BEZERRA, BYRON L. D. ; LIMA, E. B. . How Far Deep Learning Systems for Text Detection and Recognition in Natural Scenes are Affected by Occlusion?. In: ICDAR 2021 WORKSHOP ON CAMERA-BASED DOCUMENT ANALYSIS AND RECOGNITION, 2021, Lausanne. Proceedings of the 16th International Conference on Document Analysis and Recognition, 2021.

The ISTD-OC dataset is one of our contributions on "How Far Deep Learning Systems for TextDetection and Recognition in Natural Scenes are Affected by Occlusion?" published at CBDAR 2021. The dataset is derivated from a systematic methodology in which, for each image, word instances were occluded differently. The work is only a small step towards recognizing occluded texts in natural scenes robustly. Still, the proposal is sufficient for the initial analysis of current models' generalization ability in the face of occlusion.

So, to provide occluded images through the ISTD-OC dataset, we've collected 1500 images for detection and 4468 images for recognition from the ICDAR 2015 dataset. The core idea was to deliver pictures with a range of partial occlusion in a certain degree of randomness. The prediction was made considering the proportion of the text size to the original image. Therefore, occlusion corresponds to a random part of the original image overlaid on the text, and its generated by a Gaussian distribution. In this way, minimal readers are prevented from receiving very large occlusions and vice versa. The occlusion level fluctuates between 0 and 100%, intending to analyze the techniques graphically and at what occlusion performance deteriorates.

Moreover, to identify occluded texts a little more complex, we added noise to the RGB channels of the occlusion, so it is possible to prevent models from learning patterns based on equal regions to locate occlusions. Even using the ICDAR 2015 dataset as a reference to assess the occlusion problem,  once the text region is situated through the coordinates, the occlusion generation methodology allows different levels of obstruction to be applied to the text of any other desired dataset. The choice of this specific benchmark as a source is due only to its recurrent application in the analysis of baseline algorithms, avoiding possible inconsistencies during the evaluation process.

If you feel interested in the implementation code access: https://github.com/alinesoares1/ISTD-OC-Dataset. 

The collection contains offline signature samples. The signatures were collected under the supervision of Bryan Found and Doug Rogers in the years 2002 and 2006, respectively. The images were scanned at 600dpi resolution and cropped at the Netherlands Forensic Institute for the purpose of 4NSigComp2010 signature verification competition.

The dataset provide more than 11 000 expressions handwritten by hundreds of writers from different countries, merging the data sets from 4 CROHME competitions. Writers were asked to copy printed expressions from a corpus of expressions. The corpus has been designed to cover the diversity proposed by the different tasks and chosen from an existing math corpus and from expressions embedded in Wikipedia pages. Different devices have been used (different digital pen technologies, white-board input device, tablet with sensible screen), thus different scales and resolutions are used. The dataset provides only the on-line signal.

In the last competition CROHME 2013 the test part is completely original and the train part is using 5 existing data sets:

In CROHME 2014 a new test set has been created with 987 new expressions and 2 new tasks has been added: isolated symbol recognition and matrix recognition. Train and test files as the evaluation scripts for these new tasks are provided. For the isolated symbol datasets, elements are extracted from full expression using the existing datasets, which also includes segmentation errors. For the matrix recognition task, 380 new expressions have been labelled and split into training and test sets.

Furthermore, 6 participants of the 2012 competition provide their recognized expressions for the 2012 test part. This data allows research on decision fusion or evaluation metrics...

The Bentham collection consists of a set of images of a collection of works on law and moral philosophy written by the philosopher Jeremy Bentham.

The selected subset has been written by several hands (Bentham himself and his secretaries) and entails significant varibilities and difficulties regarding the quality of text images and writting styles. Training and test data were provided in the form of carefully segmented line images, along with the corresponding transcripts.

This dataset is free available for research purposes and it is provided into two parts: the images and the GT. The GT includes information about the  layout and the transcription at line level of each image in PAGE format. 

DIVA-HisDB is a collection of three medieval manuscripts that have been chosen regarding the complexity of their layout [1], together with partners from e-codices (http://www.e-codices.unifr.ch/) and the Humanities faculty in the University of Fribourg:

Preview pages below:

HANDS-VNOnDB (VNOnDB in short) provides 1,146 Vietnamese paragraphs of handwritten text composed of 7,296 lines, more than 480,000 strokes and more that 380,000 characters written by 200 Vietnamese. Writers were asked to write freely ground-truth text from a corpus of Vietnamese text. Our ground-truth text is derived from the VieTreeBank corpus which contains all of Vietnamese characters and some special symbols since it bases on Vietnamese newspapers. For collecting patterns, we used Fujistu PC Tablets (FMVT8170) with stylus pen at high sampling rate (120Hz). Each sequence contains multiple lines within various delayed strokes. VNOnDB is available for research purpose only. For commercial purposes and sell products, please contact us via email.

This is the dataset used in the Handwritten Text Recognition (HTR) contest on the READ dataset  that was held in the context of the International Conference on Frontiers in Handwriting Recognition 2016. This contest aimed to bring together researchers working on off-line HTR and provide them a suitable benchmark to compare their techniques on the task of transcribing typical historical handwritten documents. Two tracks with different conditions on the use of training data were proposed. The handwritten images for this contest were drawn from the German document Ratsprotokolle collection composed of minutes of the council meetings held from 1470 to 1805, used in the READ project. The selected dataset has been written by several hands and entails significant variabilities and difficulties.

The dataset provides more than 12 000 expressions handwritten by hundreds of writers from different countries, merging the data sets from 4 previous CROHME competitions and adding new ressources. Writers were asked to copy printed expressions from a corpus of expressions. The corpus has been designed to cover the diversity proposed by the different tasks and chosen from an existing math corpus and from expressions embedded in Wikipedia pages. Different devices have been used (different digital pen technologies, white-board input device, tablet with sensible screen), thus different scales and resolutions are used. The dataset provides only the on-line signal.

The training and validation sets are using CROHME 2014 datasets, but 1147 new on-line handwritten expressions are used in the CROHME 2016 test dataset. Furthemore, this new package provide updated version of tools and new math ressources: more than 500K of LaTeX expressions from wikipedia to train language models.

All these ressources allow to adress 4 different tasks: Formulas recognition, Structure recognition, isolated Symbol recognition and Matrices recognition.

There are 5,400 Thai signatures in this dataset, 3,000 (from 100 signers) are genuine signatures, for each genuine signature, there are 1,200 skillfully forged and 1,200 simple forged signatures.

There are 8,400 name components (first and last names) from 100 students in this dataset. There are 6,000 genuine name components and 2,400 (100 students × 2 name components × 12 times) skilfully forged name components. 

HANDS-VNOnDB2018 (VNOnDB2018 in short [1]) is used for ICFHR2018 Competition on Vietnamese Online Handwritten Text Recognition using VNOnDB. It provides 1,146 Vietnamese paragraphs of handwritten text composed of 7,296 lines, more than 480,000 strokes and more that 380,000 characters written by 200 Vietnamese.

Writers were asked to write freely ground-truth text from a corpus of Vietnamese text. Our ground-truth text is derived from the VieTreeBank corpus which contains all of Vietnamese characters and some special symbols since it bases on Vietnamese newspapers. For collecting patterns, we used Fujistu PC Tablets (FMVT8170) with stylus pen at high sampling rate (120Hz). Each sequence contains multiple lines within various delayed strokes. VNOnDB2018 is available for research purpose only. For commercial purposes and sell products, please contact us via email.

The dataset contains 40 images for training and validation and 10 images for testing.

The document pages in the dataset are from multiple sources which are digitized using different devices. This increased variance makes the dataset especially challenging for segmentation task.