This repo contains code to extend/replicate the dataset present in the Kaggle Bengali.AI Handwritten Grapheme Classification. For the dataset, codes, discussions and leaderboards, visit the Kaggle competition page.

The paper with added appendix can be found here.

Common Handwritten Graphemes in Context

Project Structure

.
- data
   -- scanned
   -- extracted
   -- error
   -- packed
- codes
- collection
   -- A4
   -- Letter
- logs

Basic Usage

1. Run python ./data/extracted/purge.py to clear extraction folders
2. Download and extract batch of scanned file .jpgs to ./data/scanned/<batchname>
3. cd ./data/scanned and run python transcribeGui.py <batchname>
4. After Roll/ID are transcribed execute extract.m on MATLAB. Specify source folder before executing. Replace surfAlignGPU() with surfAlign in the absence of GPU support. Set disp=true for ocrForm(), surfAlign(), surfAlignGPU() to validate extraction performance. For surfAlign() set nonrigid=true.
5. cd ./data/error and check for extraction failures.
6. cd ./data/extracted and check for label errors in sub-folders.
7. Run python pack.py which will create separate folders for each extracted <batchname> inside ./data/packed.
8. cd ./data/packed/ and run python labelXGui.py <batchname>. Select overwriting and empty blobs to be discarded and Ctrl+S to save. After you are done going through all of the packets, click the transfer button to remove errors from the packaged folder.

Dependencies

• MATLAB 2017b or higher

• MATLAB Computer Vision Toolbox

• Python 3.6.3 or higher

• Pillow == 4.2.1

Documentation

• Kaggle competition page www.kaggle.com/c/bengaliai-cv19
• Dataset introduction COCO-Grapheme