View source

{{Dataset <!-- please remove unassigned parameters -->
| domain =  Computer Vision   <!-- e.g. Medical, News, Picture -->
| media =   Image   <!-- e.g. Image, Music, Video, Text -->
| size =       <!-- in MB, GB -->
| instances = 6000 <!-- number of instances -->
| format =    JPEG <!-- file format (e.g. TXT, PNG, MP3, ...) -->
| created =    <!-- the creation date: YYYY-MM-DD -->
| task =     Pattern Recognition  <!-- the typical task: e.g. Retrieval, Classification, Recognition -->
| copyright =  <!-- copyright remarks, please put details below -->
| url =    http://dag.cvc.uab.es/MIPRCV_benchmarks/CD_COVERS/    <!-- where to get this dataset -->
}}

== Description ==

This dataset is composed of 6.000 CD/DVD cover images and some associated labels. The term “cover” refers to the font-facing panel of a CD/DVD package, and, increasingly, the primary image accompanying a digital download of the album, or of its individual tracks.
These images have been downloaded from amazon.com using a Java custom application making use of the Amazon API (www.amazon.com/webservices).

CD covers represent an interesting challenge related to several computer vision and pattern recognition problems. In the present state of the dataset, labels are related to regions of the image where we can find printed text and it is designed for studying the problem of unconstrained text detection in complex backgrounds.

== Quality ==



== Source ==

This benchmark is being organized by the following members of the PR-CVC group: Joan Alabort (he compiled the images and created the XML files), Xavier Baró, Sergio Escalera, Camp Davesa, Jordi Vitrià.

The building of the CD COVER database has been partially supported by MICINN Grant TIN2009-14404-C02-00 and by CONSOLIDER-INGENIO 2010 Grant CSD2007-00018.


== Ground Truth Annotation ==

Labels are stored in .xml files.                                        
Original images are identified by a unique name, composed of the name of the subset of images (training, validation, test) and a unique consecutive sequence number, for example: training_0001.jpg. 

For each image file (training_0001.jpg) there exists a xml file (training_0001.xml) with the following information: the name of the image, text regions (defined by their rectangular bounding box vertices), artist, title of the CD and original URL of the image.


== Features ==




== Copyright Remarks ==



== Citation ==

If you are using this database, please Please, cite as: 

S. Escalera, X. Baró, J. Vitrià, and P. Radeva, Text Detection in Urban Scenes, In Frontiers in Artificial Intelligence and Applications, Vol. 202, Pages 35-44,  2009, ISBN 0922-6389, IOS Press  Amsterdam, The Netherlands.


== External Links ==

http://dag.cvc.uab.es/MIPRCV_benchmarks/CD_COVERS/