Pytorch code to train scene text recognition models for Indic languages
Run the following commands if the data is in MLT19 format. If the data is already in lmdb format skip to Data Setup.
python prepare_data.py --image_dir path/to/SceneImages --image_gt_dir path/to/SceneImage_gt --word_image_dir path/to/store/cropped/text/images --output_path path/to/store/lmdb_gt --lang language to generate gt for
python create_lmdbdataset.py --gtFile path/to/lmbd_gt --outputPath folder/to/store/lmdb_data
- To get the train-test split, run
python train_test_split_lmdb.py --lmdb_data_path path/to/lmdb_data
-
Set up data directories as specified in this file.
-
Synthetic data for several Indic languages, in the desired format, can be obtained from here.
- Refer to Config.py on how to setup a config for an experiment run
- To start training run
python train.py --task_config_name <name of the config>
To add a new language head append character set of new language to the characters.txt file.
| Language | Word Accuracy (%) | N-ED (%) | Model |
|---|---|---|---|
| Hindi(hin) | 45.00 | 65.50 | gdrive |
| Bangla(ban) | 74.00 | 87.50 | gdrive |
| Kanda(kan) | 56.00 | 77.50 | gdrive |
| Malayalam(mal) | 61.00 | 75.00 | gdrive |
| Marathi(mar) | 50.50 | 72.50 | gdrive |
| odia(odia) | 37.50 | 61.20 | gdrive |
| Sanskrit(sans) | 82.00 | 91.00 | gdrive |
| Tamil(tam) | 58.00 | 77.00 | gdrive |
| Telugu(tel) | 51.00 | 80.00 | gdrive |
| Language | Word Accuracy (%) | N-ED (%) |
|---|---|---|
| Hindi(hin) | 47.80 | 86.00 |
| Bangla(ban) | 45.80 | 78.00 |