ICDAR 2017 Competition on Baseline Detection (cBAD)

The database of Track A [Simple Documents] consists of 755 images extracted from 9 different archival collections. The dataset comprises images with additional PAGE XMLs [1]. The PAGE XML contains text regions, e.g. paragraphs. Thus a layout analysis or text detection needs not to be performed on this dataset. Only handwritten text is present and the dataset contains no tables. The groundtruth of the test-set will be released after evaluating all submitted methods and the final results being made public.

Track B [Complex Documents] contains mixed documents. Though most documents are handwritten, printed documents, book covers, empty pages, and tables are contained in this track. While Track A has locally skewed text-lines, text-lines in Track B are rotated up to 180°.

[1] PAGE XML documentation with reference implementations
[2] Competition description


Track A [Simple Documents] - Download Training Set | Download Test Set
Track A [Complex Documents] - Download Training Set | Download Test Set

Evaluation Scheme

A new evaluation scheme is introduced that measures errors using baselines. A detailed description of the evaluation scheme is available here. The evaluation tool which will be used for the competition is available as standalone jar .

Submission Protocol

The submitted result file has to be a compressed tar file (.tar.gz), containing separate result files (imagename.txt or imagename.xml) for each image in the test set. The folder structure must not be changed! If the test-set contains 3 different images (in 2 subfolders):
  • /a/1.jpg
  • /a/2.jpg
  • /b/1.jpg
than the corresponding result file should contain:
  • /a/1.jpg.txt
  • /a/2.jpg.txt
  • /b/1.jpg.txt
Do not include the images in the result file. Even if your algorithm doesn't detect any text line on a certain image, you have to provide an "empty" result file for this image. A result file for a certain image could either be a valid PAGE xml-file containing the baselines detected by your method, or a txt-file containing the detected baselines. If you use text files, the baselines need to be encoded as follows:
Each row corresponds to a baseline. Different points are semicolon separated, x- and y-coordinates are comma-separated: x1,y1;x2,y2;x3,y3.

Subtrack 1.2 - Track B [Complex Documents]


14th June 2017

Submission deadline extended

24th May 2017

Training sets updated

There is a second competition on baseline detection

01st Mar. 2017

Datasets are online

17st Jan. 2017

Competition site is online

Important Dates

Registration Deadline

18th June 2017

Submission Deadline

07th July 2017

ICDAR 2017

10th-15th Nov. 2017


Markus Diem

[TU Wien, Computer Vision Lab] 

Markus Diem is a senior scientist at the Computer Vision Lab, TU Wien, Austria. His research interests are Cultural Heritage Applications and Document Analysis.

Florian Kleber

[TU Wien, Computer Vision Lab] 

Florian Kleber is currently a senior scientist at the Computer Vision Lab, Institute for Computer Aided Automation, TU Wien, Austria. His research interests are Cultural Heritage Applications and Document Analysis Applications.

Stefan Fiel

[TU Wien, Computer Vision Lab] 

Stefan Fiel works as research assistant at the Computer Vision Lab, TU Wien.

Basilis Gatos

[NCSR Demokritos / IIT / CIL] 

Researcher at the Institute of Informatics and Telecommunications of the National Center for Scientific Research "Demokritos", Athens, Greece.