CVL OCR DB
(Computer Vision Lab OCR DataBase)

A public annotated image dataset of text in natural scenes

  
  
  

About CVL OCR DB

CVL OCR DB is a public annotated image dataset of text in natural scenes. Images include signboards, shop names, traffic signs, jumbo posters etc. CVL OCR DB is suitable for evaluating both text detection and text recognition methods. Two types of annotation are supported: binary annotation and polygon annotation. Dataset includes additional cropped single character images suitable for OCR training and testing.

Features:

  • 120 binary annotated (text/non-text) images of text in natural scenes (NEW!!!)
  • 341 polygon annotated images of text in natural scenes
  • 7014 cropped single character images suitable for OCR training and testing
  • all images are in JPEG image format
  • includes TextAnnotator - a powerful annotation software for annotating your own images
  • publicly available free of charge

Authors: Andrej Ikica & Peter Peer

Current version: 1.3

Binary annotation (NEW!!!)

Each binary annotated image in CVL OCR DB dataset has a corresponding binary (text/non-text) BMP image (see figure below).

(a)
(b)

Figure: An example of binary annotation.
(a) Original JPEG image. (b) Binary annotated BMP image.

Polygon annotation & cropped character images

Instead of annotating text regions with rectangles all text regions in CVL OCR DB images are annotated with poylgons (see figure below). Each polygon annotated image in CVL OCR DB has a corresponding XML annotation file (see Xml annotation file for details). Additionally, images of cropped characters are available for each CVL OCR DB image, thus making CVL OCR DB also suitable for OCR training and testing.

(a)
(b)
(c)

Figure: An example of polygon annotation.
(a) Original JPEG image. (b) Text region annotated with n-polygon. (c) Cropped characters.

Material & documentation

Material & documentation

Additional documentation:

Obtaining CVL OCR DB

Terms of usage:

  • license agreement must be signed
  • non-commercial usage
  • research in computer vision
  • a proper citation must be included in your publications

Contact:

Send the scan of the filled and hand signed license agreement to andrej.ikica@fri.uni-lj.si in order to obtain the CVL OCR DB dataset.


 
Andrej Ikica, 2013