Show, Attend and Read - A PyTorch Implementation

Implementation of Show, Attend and Read: A Simple and Strong Baseline for Irregular Text Recognition in AAAI 2019, with PyTorch >= v1.4.0.

Task

• Backbone model
• Encoder model
• Decoder model
• Integrated model
• Data processing
• Training pipeline
• Inference pipeline

Supported Dataset

• Street View Text: http://vision.ucsd.edu/~kai/svt/
• IIIT5K: https://cvit.iiit.ac.in/research/projects/cvit-projects/the-iiit-5k-word-dataset
• Syn90k: https://www.robots.ox.ac.uk/~vgg/data/text/
• SynthText: https://www.robots.ox.ac.uk/~vgg/data/scenetext/

Command

Training

python train.py --batch 32 --epoch 5000 --dataset ./svt --dataset_type svt --gpu True

Inference

python inference.py --batch 32 --input input_folder --model model_path --gpu True

Results

SVT

IIIT5K

Input:

Output attention map per character:

Syn90K (10k for training/3k for testing)

Input:

Output attention map per character:

SynthText (80k for training/20k for testing)

Input:

Output attention map per character:

Source

[1] Original paper: https://arxiv.org/abs/1811.00751

[2] Official code by the authors in torch: https://github.com/wangpengnorman/SAR-Strong-Baseline-for-Text-Recognition

[3] A TensorFlow implementation: https://github.com/Pay20Y/SAR_TF