EAST Scene Text Detector with MobileNetV3

This is a Pytorch re-implementation of EAST: An Efficient and Accurate Scene Text Detector using MobileNetV3 as the feature extractor.
Most parts of this implementation are taken from https://github.com/argman/EAST

Prerequisites

This project using Pytorch - An open source machine learning framework

Installing

Download the project

git clone https://github.com/ishin-pie/east-mobilenet.git

Installing from requirements.txt file

cd east-mobilenet

pip install -r requirements.txt

Note: we suggest you to install on the python virtual environment
Learn more: Installing Deep Learning Frameworks on Ubuntu with CUDA support

Running Demo

Running the demo of our pre-trained model

python demo.py -m=model_best.pth.tar -i=[path to your image]

checking your result in "demo" folder

or you can run camera demo (work fine on GPU machine)

python camera_demo.py -m=model_best.pth.tar

Dataset Structure

During training, we use ICDAR 2015 Traning Set and ICDAR 2017 Training Set (Latin only). In addition, we use ICDAR 2015 Test Set for validating our model.

The dataset should structure as follows:

[dataset root directory]
├── train_images
│   ├── img_1.jpg
│   ├── img_2.jpg
│   └── ...
├── train_gts
│   ├── gt_img_1.txt
│   ├── gt_img_2.txt
│   └── ...
└── test_images
    ├── img_1.jpg
    ├── img_2.jpg
    └── ...

Note: the [dataset root directory] should be placed in "config.json" file.

Sample of ground truth format:

x1,y1,x2,y2,x3,y3,x4,y4,script,transcription

Training

Training the model by yourself

python train.py

Note: check the "config.json" file, which is used to adjust the training configuration.

Experiment on GEFORCE RTX 2070

Loss

Examples

Acknowledgments

• https://github.com/argman/EAST
• https://github.com/jiangxiluning/FOTS.PyTorch