This python script will generate n pages of words with bbox and its ground truth labels. Also it supports various background colors, fonts etc. Additionally it can export the dataset as tfrecord
The code is tested and developed in Ubuntu 20.04 and using Pyton 3.8.But the code has the realiability to run on most of the configuration . If you face issues , do open up an issue for this repo .All the package dependencies are mentioned in requirements.txt.
For Word Generator
------------------
--output_dir: The datset images to be stored (default: dataset/)
--input_file: Text file contain random words for generator dataset pages
--background: Background Color (default: white)
--font_dir: Fonts to be used for generating dataset (default: fonts/)
--num_pages: Number of images of dataset need to be generated (default: 10)
--width: Width of the image (default: 600) in pixel
--height: Height of the image (default: 800) in pixel
For TFRecord Generator
----------------------
--csv_input: Ground truth labels csv file (default: ground_truth.csv)
--output_path: Location for tfrecord file need to be saved (default: dataset.tfrecord)
--dataset_dir: Dataset dir need to be used for images (default: dataset/)
- Install python 3.8 and requirements.txt to install the necessary dependencies
- To run the word detection generator.
python insert_word.py --input_file words.txt --num_pages 100 - The dataset will be stored in
datasetfolder and the coordinates, ground truth values will be save inground_truth.csv - To export as tfrecord file,
python generate_tfrecord.py - To check the bbox drawn in image use
cv_doc.py - Enjoy the dataset
Copyright © 2021 Robin Reni. All rights reserved