EASTLite is a lightweight implementation of EAST based on MobileNet, which significantly speeds up the execution and makes the algorithm friendlier to resource-constrained devices. It is written with TensorFlow 2 and Keras's Functional API.

Credit to https://github.com/argman/EAST, whose code we have used for geometric operations and data augmentation in the ground truth generations, and for the post-processing and NMS. You can also change the base model to ResNet by using the flag --base_model resnet.

You can very easily set up a demo web by building and running the Dockerfile. This is the easiest way to set up GPU acceleration easily using NVIDIA Container Toolkit.

Build the image with

docker build -t east-lite-web .

and run a GPU-accelerated container with

docker run --name east-lite-web -p 8000:8000 -p 3000:3000 -d --gpus all east-lite-web

Then navigate to http://localhost:3000 and check out the algorithm and its speed! You can find out whether GPU acceleration is on via the Docker logs (with docker logs east-lite-web) or by GETting the backend's root at http://localhost:8000.

Download the pretrained weights. This model obtains an F-score of 0.749 on ICDAR15 (if you train a better version please submit it 😊).

After downloading the checkpoint, you can run inference by executing

  python infer.py --checkpoint_path /tmp/ckpt --data_path demo_inputs/ --visualize --dont_write

You can train the network by passing the following arguments with the path to the training and validation sets:

  python train.py --training_data_path /tmp/train_ds/ --validation_data_path /tmp/val_ds/ --batch_size B

For a full list of flags run python train.py --help.

The dataset must be comprised of pictures and txt files in the format of the ICDAR15 ground truth text files.

158,128,411,128,411,181,158,181,Footpath
443,128,501,128,501,169,443,169,To
64,200,363,200,363,243,64,243,Colchester
394,199,487,199,487,239,394,239,and
72,271,382,271,382,312,72,312,Greenstead  

In our training process, we have used ICDAR15's and ICDAR13's training datasets for training, and ICDAR15's test set for validation.