lipnet-pytorch's Introduction

LipNet: End-to-End Sentence-level Lipreading

The state-of-art PyTorch implementation of 'LipNet: End-to-End Sentence-level Lipreading' by Yannis M. Assael, Brendan Shillingford, Shimon Whiteson, and Nando de Freitas (https://arxiv.org/abs/1611.01599). This version achieves the best performance in all evaluation metrics.

Results

Scenario	Image Size (W x H)	CER	WER
Unseen speakers (Origin)	100 x 50	6.7%	13.6%
Overlapped speakers (Origin)	100 x 50	2.0%	5.6%
Unseen speakers (Ours)	128 x 64	6.7%	13.3%
Overlapped speakers (Ours)	128 x 64	1.9%	4.6%

Notes:

Contribution in sharing the results of this model is highly appreciated

Data Statistics

Scenario	Train	Validation
Unseen speakers (Origin)	28775	3971
Overlapped speakers (Origin)	24331	8415
Unseen speakers (Ours)	28837	3986
Overlapped speakers (Ours)	24408	8415

Preprocessing

Link of processed lip images and text: https://pan.baidu.com/s/165swxO8A-GUZ03-InTL6VA

Google Drive: https://drive.google.com/drive/folders/1Wn2EJw2101nF59eNDXEto6qXqfgDDucL

Download all parts and concatenate the files using the command

cat GRID_LIP_160x80_TXT.zip.* > GRID_LIP_160x80_TXT.zip
unzip GRID_LIP_160x80_TXT.zip
rm GRID_LIP_160x80_TXT.zip

We provide examples of face detection and alignment in scripts/ folder for your own dataset.

Training And Testing

python main.py

Data path and hyperparameters are configured in options.py. Please pay attention that you may need to modify options.py to make the program work as expected.

Dependencies

PyTorch 1.0+
opencv-python

Bibtex

@article{assael2016lipnet,
  title={LipNet: End-to-End Sentence-level Lipreading},
  author={Assael, Yannis M and Shillingford, Brendan and Whiteson, Shimon and de Freitas, Nando},
  journal={GPU Technology Conference},
  year={2017},
  url={https://github.com/Fengdalu/LipNet-PyTorch}
}