To build the project, first install the latest versions of Torch7 and LMDBOn Ubuntu, lmdb can be installed by apt-get install liblmdb-dev. The following project also needs to be installed: https://github.com/qassemoquab/stnbhwd

Instructions for Testing

1. Create a text file containig path to list of images for which textual recognition is required
2. Run th src/test.lua (If you want to specifiy parent path, you can set the addn variable)
3. If you are unable to run the code, you have to install fblualib and then go to src/ and execute sh build_cpp.sh to build the C++ code. If successful, a file named libcrnn.so should be produced in the src/ directory.
4. If you wish to do lexicon based decoding, you need to create a lexion file and that needs to be present in the src/ folder. The lexicon file would contain, say the set of unique words in the test set. You can change the name of the lexicon file in the test.lua script. (We don't do test time augmentation as done in the paper to save time. Also the WER/CER calculation done in the code here is Approximate.)

Instructions for Training -- Following the stucture in IIIT-HW-Dev dataset, one would have train/test/val files with image-path and label pairs

1. Depending on whether you want to create a dataset with/without augmented images, run tools/create_dataset or tools/create_dataset_aug respectively. Either class requires outputPath to where the lmdb file is to be written, imagePathfileLilst which is just the train/test/val .txt file from the dataset, parentDirofImages, which mentions the directory where the HindiSeg folder is present and the lookupFile which was created in step 2.
2. Go to the models/ folder and either create new folder or copy an existing folder. The various properties present that need to be changed in config.lua file are mentioend below.
3. Run the training code as: src/th main_train.lua <Path to snapshot / saved model> ( <Path to snapshot / saved model> is optional.)

PS: The code is based upon the work done here: https://github.com/bgshih/crnn (Search in that repos forum in case of doubts) IAM dataset link: http://www.fki.inf.unibe.ch/databases/iam-handwriting-database

A basisc description of the various files and folders Model Folder: config.lua:: Contains all the various arch. defn's like number of classes and the layers.

snapshotInterval: After how many iter. a model is saved to disk

nClasses: no of unique char's in the dataset. maxT: See from the model defn Model cannot recognize words having more characters than maxT.

savePath: Path where the various snapshots are saved testInterval: After how many iterations you get test results trainSetPath and valSetPath need to be set where the training and validation lmdb files are present

In src folder: DatasetLmdb.lua --> only need to change local imgW, imgH everywhere

test.lua --> use for predicting results on test set, change snapshot path and wether to use lexicon free or lexicon based decoding.

inference.lua --> lexicon size is fixed to 30 characters.

utilities.lua ascii2label: ranking always starts from 1 ends at nClasses in config.lua

label2ascii: opposite of ascii2label function and be careful of label ==0 case

loadAndResizeImage: Needs to be the set to the same values in DatasetLmdb.lua