Summary

Images of Kannada characters cropped from street photos

License

ODbL

Dependencies

Tags

character-recognition computer-vision Kannada transfer-learning

Attribute Types

Download

tgz (106.0 MB)

Files are converted on demand and the process can take up to a minute. Please wait until download begins.

Original Data Format

tgz

Name

Version mldata

Comment

Names

Data (first 10 data points)

Description

This dataset contains images (bitmaps in PNG format) of symbols used in the Kannada language. Kannada symbols are phonetic, so consonant symbols are combined with vowel symbols to for syllable symbols. The syllable symbols (compound) were treated as individual classes, i.e., our annotation is flat. This means that when two symbols are combined, a third symbol is generated, which can lead to the creation of almost 1000 classes. This dataset contains a subset of 563 classes some of the classes have a very small number of samples. These features make the classification problem very challenging. A finer annotation is required, indicating how compound symbols are generated. Methods that take advantage of this should lead to much better recognition rates than the ones reported in our paper.

The TGZ file contains two sub-directories: GoodImg and BadImg. The images in the GoodImg directory have low level of occlusion and noise, and their labels have been verified by a native speaker. In [deCampos et al. VISAPP2009], we did not use the set in BadImg. Each of these directories, has a Bmp and a Msk sub-directory, containing image bitmaps and the binary segmentation masks, respectively. For most of the images, the masks are rectangular, but often the masks are helpful to exclude parts of neighbouring characters. Each subdirectory is in the format SampleXXX, where XXX is the class label.

Acknowledgements

This dataset and the experiments present in the paper were done at Microsoft Research India by Teo deCampos, with the mentoring support from Manik Varma. Additional SVM and MKL experiments were performed by Bodla Rakesh Babu.

We would like to acknowledge the help of several volunteers who annotated this dataset. In particular, we would like to thank Arun, Kavya, Ranjeetha, Riaz and Yuvraj. We would also like to thank Richa Singh and Gopal Srinivasa for developing some of the tools for annotation (one of the tools used is described here). We are grateful to CV Jawahar for helpful discussions.

We kindly request that any publication obtained by using this dataset cites our original benchmark paper [deCampos et al VISAPP2009].

URLs

http://www.ee.surrey.ac.uk/CVSSP/demos/chars74k/

Publications

Data Source

This dataset was generated by manually segmenting out characters that appear in sign board in public areas of Bangalore and other places in Karnataka. The original images and annotations in Unicode can be obtained from http://www.ee.surrey.ac.uk/CVSSP/demos/chars74k/.

Measurement Details

These are raw bitmap images. In [deCampos et al. VISAPP2009] we evaluated several feature extraction methods in a bag-of-visual-words framework.

Binary segmentation maps are also available.

Usage Scenario

This dataset was created to evaluate camera-based character recognition techniques. It can also be used for transductive transfer learning methods, e.g. if the Kannada hnd dataset is used as source domain and this dataset is used as target.

revision 1

by teo on 2012-03-27 16:16

revision 2

by teo on 2012-09-24 21:22

revision 3

by teo on 2012-09-24 21:55

revision 4

by teo on 2012-09-24 21:56

revision 5

by teo on 2012-09-24 21:59

revision 6

by teo on 2012-09-24 22:09

revision 7

by teo on 2012-09-24 22:11

revision 8

by teo on 2012-09-24 22:12

revision 9

by teo on 2012-09-24 22:14

revision 10

by teo on 2012-09-24 22:20

No one has posted any comments yet. Perhaps you would like to be the first?

Leave a comment

This item was downloaded 3027 times and viewed 12967 times.

No Tasks yet on dataset Chars74K Kannada img