8

I'm looking for a large labelled data set, of characters that I can use to contrast recognising Letters, with recognising Numbers.

Something very similar to MNIST, but with letters. Language doesn't matter, other than it must be simpler than Japanese Kanji. (<500 distinct characters, not extremely visually complex)

Ideally 28x28 greyscale, with the letters taking up a 20x20 box.

5
  • Some of the various collaborative transcription projects (eg, Notes from Nature, Ancient Lives, Old Weather might be able to be used as input ... although I think only with Ancient Lives do you denote location + character individually. (and it's Greek).
    – Joe
    Dec 26, 2013 at 17:53
  • It being Greek is Ok, but it is explictly no a labelled dataset. Dec 27, 2013 at 0:30
  • it may not be now, and you'd likely have to do some cleanup work to get it to your preferred dimensions .. but you can try contacting the people running the project and see if they can give you what you'd need to create your training set.
    – Joe
    Dec 27, 2013 at 15:49
  • @Oxinabox Why do you need this? (I'm just curious.) Feb 19, 2015 at 7:28
  • This was 12 months ago, but I was preparing for a research project about knowledge transfer between deep neural networks. I have results from that project now, but there is a limitation in scope that I only realized after it was complete that means it is not particularly publishable. (It only considers single epoch of training in its comparisons). One day I may do a follow up and cover the missing areas. Feb 19, 2015 at 7:54

3 Answers 3

4

I created a dataset with of on-line data. It has 369 symbols (including a-z A-Z 0-9 \alpha-\omega), but it is online data. You will have to create the rendered versions yourself:

http://write-math.com/data

Each class has at least 50 recordings. (I don't have much data for letters, so it will probably not be more.)

edit: I've redered it. See The HASYv2 dataset.

3

This page lists some on/off-line handwriting database for academic use. Some of them can be downloaded free while others may need application.

http://www.gavo.t.u-tokyo.ac.jp/~qiao/database.html

1
  • I ran into this same list. However, many links are bad on this list. The CEDAR one costs. The cambridge one may be found here svr-ftp.eng.cam.ac.uk/pub/data/lob.tar Hopefully someone other than myself found better resources than this.
    – brianray
    Apr 21, 2014 at 19:03
1

Do check an image dataset at KAGGLE it contains A-Z handwritten 370000+ images

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.