This dataset contains:
1. Cropped word images split into training and test sets
2. Ground truth annotation, small and medium sized lexicons
3. Lexicon with 0.5 million words (from Weinman et al. 2009)
4. Character bounding box level annotations
5. Easy-hard word splits (Please mail us to obtain this)

The lexicon used to compute language priors is in the file lexicon.txt. Please
use this lexicon when comparing with our large size lexicon based recognition
results. This lexicon was provided by Weinman et al. 2009. The following paper
should be cited when using this lexicon.

@article{Weinman09,
    author = {Jerod J. Weinman and Erik Learned-Miller and Allen Hanson},
    title  = {Scene Text Recognition using Similarity and a Lexicon with Sparse Belief Propagation},
    journal= {IEEE Trans. Pattern Analysis and Machine Intelligence},
    volume = {31},
    number = {10},
    pages  = {1733--1746},
    month  = {Oct},
    year   = {2009}
}


How to load test data information
---------------------------------
(Usage: Case insensitive small/medium/large lexicon cropped word recognition)

1. Open Matlab
2. Load testdata
3. A structure testdata will be loaded. This structure has four fields.
   (a) ImgName       The cropped word image name.

   (b) GroundTruth   Specifies the ground truth text corresponding to the 
                     cropped word
   (c) smallLexi     Contains a lexicon list of 50 words per image (referred to
                     as small size lexicon in the paper)
   (d) mediumLexi    Contains a lexicon list of 1000 words per image (the
                     medium size lexicon) 


How to load character bounding box information
-----------------------------------------------------------------
(Usage: Case sensitive character detection/recognition)

1. Open Matlab
2. Load testCharBound (or trainCharBound)
3. A structure testCharBound (or trainCharBound) will be loaded. It contains
   three fields.
   (a) ImgName  The word image name
   (b) chars    A string of characters.
   (c) charBB   Bounding boxes of characters in same order as chars. Bounding box
                are stored as [x y width height].
  

Example:
>> load testCharBound
 
>> testCharBound(1).ImgName

ans =

test/1002_1.png


>> testCharBound(1).chars

ans =

PRIVATE


>> testCharBound(1).charBB(1,:) %% Loads bounding box for character "P" (i.e. first character of testCharBound(1).chars)

ans =

     4     7    32    45

>> testCharBound(1).charBB(5,:) %% Loads bounding box for character "A" (i.e. fifth character of testCharBound(1).chars)

ans =

   115     7    37    43


-------------------------------------------------------------------------------
If you use this dataset, please cite the following paper.

@InProceedings{MishraBMVC12,
 author   = "Mishra, A. and Alahari, K. and Jawahar, C.~V.",
 title    = "Scene Text Recognition using Higher Order Language Priors",
 booktitle= "BMVC",
 year     = "2012"
}

For any queries about the dataset contact: anand.mishra@research.iiit.ac.in