Media-Text

Media-Text dataset comprising images of banners, posters, covers and another images characterised for media industry.

Full paper is available here: Media-Text: a Media Industry-Based Dataset for Scene Text Detection

DATASET DESCRIPTION

• 400 images
• 7 744 annotated text instances
• 973 annotations have been marked as illegible for the task of text recognition
• 659 texts have been markes as do not care (###) for scene text detection.
• Images are represented by 193 unique resolutions.

Annotation Format - Each image has corresponding  gt_*.txt file, which contains annotations in bounding box format (defined by 4 courners), transcription, and bool flag which determines that text is illegible for OCR. Proposed format is similar to ICDAR15 annotations.

x1, x2, ..., x4, y4, transcription, OCR Flag 

Example:

37,68,198,49,214,181,52,200,LADIES,False

ACKNOWLEDGMENT

This work was supported by the Silesian University of Technology (SUT) through the subsidy for maintaining and developing research potential grant in 2024 for young researchers, No. 2/070/BKM24/0058, and by the Ministry of Science and Higher Education "Implementation Doctorate" No. DWD/5/0511/2021.

Thanks to the graphic department of media-press group for the preparation and possibility of sharing graphics thematically related to the prepared dataset.

LICENSE

Annotations created by authors are licesned under CC-BY-4.0 license.Images from the Open-Image-V7 dataset and are licensed according to their source information. Source information is defined in a file metadata.csv file that defines all the metadata of each file (File name corresponds to the ImageID column).

Images whose name corresponds to the media_press pattern are provided for academic use.