Multi-modality Associative Bridging through Memory.
Application in Lip Reading : Cross-modal Memory Augmented Visual Speech Recognition

This repository contains the official PyTorch implementation of the following paper:

Multi-modality Associative Bridging through Memory: Speech Sound Recollected from Face
Minsu Kim, Joanna Hong, Sejin Park, and Yong Man Ro
Paper: https://openaccess.thecvf.com/content/ICCV2021/papers/Kim_Multi-Modality_Associative_Bridging_Through_Memory_Speech_Sound_Recollected_From_Face_ICCV_2021_paper.pdf

Preparation

Requirements

• python 3.7
• pytorch 1.6 ~ 1.9
• torchvision
• torchaudio
• av
• tensorboard
• pillow

Datasets

LRW dataset can be downloaded from the below link.

• https://www.robots.ox.ac.uk/~vgg/data/lip_reading/lrw1.html

The pre-processing will be done in the data loader.
The video is cropped with the bounding box [x1:59, y1:95, x2:195, y2:231].

Training the Model

main.py saves the weights in --checkpoint_dir and shows the training logs in ./runs.

To train the model, run following command:

# Distributed training example for LRW
python -m torch.distributed.launch --nproc_per_node='number of gpus' main.py \
--lrw 'enter_data_path' \
--checkpoint_dir 'enter_the_path_for_save' \
--batch_size 80 --epochs 200 \
--mode train --radius 16 --n_slot 88 \
--augmentations True --distributed True --dataparallel False\
--gpu 0,1...

# Data Parallel training example for LRW
python main.py \
--lrw 'enter_data_path' \
--checkpoint_dir 'enter_the_path_for_save' \
--batch_size 320 --epochs 200 \
--mode train --radius 16 --n_slot 88 \
--augmentations True --distributed False --dataparallel True \
--gpu 0,1...

Descriptions of training parameters are as follows:

• --lrw: training dataset location (lrw)
• --checkpoint_dir: directory for saving checkpoints
• --batch_size: batch size --epochs: number of epochs --mode: train / val / test
• --augmentations: whether performing augmentation --distributed: Use DataDistributedParallel --dataparallel: Use DataParallel
• --gpu: gpu for using --lr: learning rate --n_slot: memory slot size --radius: scaling factor for addressing score
• Refer to train.py for the other training parameters

Testing the Model

To test the model, run following command:

# Testing example for LRW
python main.py \
--lrw 'enter_data_path' \
--checkpoint 'enter_the_checkpoint_path' \
--batch_size 80 \
--mode test --radius 16 --n_slot 88 \
--test_aug True --distributed False --dataparallel False \
--gpu 0

Descriptions of training parameters are as follows:

• --lrw: training dataset location (lrw)
• --checkpoint: the checkpoint file
• --batch_size: batch size --mode: train / val / test
• --test_aug: whether performing test time augmentation --distributed: Use DataDistributedParallel --dataparallel: Use DataParallel
• --gpu: gpu for using --lr: learning rate --n_slot: memory slot size --radius: scaling factor for addressing score
• Refer to test.py for the other testing parameters

Pretrained Models

You can download the pretrained models.
Put the ckpt in './data/'

Bi-GRU Backend

• https://drive.google.com/file/d/1wkgkRWxu7JM0uaNHmcyCpvVz9OFar8Do/view?usp=sharing
  

To test the pretrained model, run following command:

# Testing example for LRW
python main.py \
--lrw 'enter_data_path' \
--checkpoint ./data/GRU_Back_Ckpt.ckpt \
--batch_size 80 --backend GRU\
--mode test --radius 16 --n_slot 88 \
--test_aug True --distributed False --dataparallel False \
--gpu 0

MS-TCN Backend

• https://drive.google.com/file/d/1uHZbmk9fgMqYVfvaoMUe-9XlGvQnXEcS/view?usp=sharing

To test the pretrained model, run following command:

# Testing example for LRW
python main.py \
--lrw 'enter_data_path' \
--checkpoint ./data/MSTCN_Back_Ckpt.ckpt \
--batch_size 80 --backend MSTCN\
--mode test --radius 16 --n_slot 168 \
--test_aug True --distributed False --dataparallel False \
--gpu 0

Architecture	Acc.
Resnet18 + MS-TCN + Multi-modal Mem	85.408
Resnet18 + Bi-GRU + Multi-modal Mem	85.864

AVSR

You can also use the pre-trained model to perform Audio Visual Speech Recognition (AVSR), since it is trained with both audio and video inputs.
In order to use AVSR, just use ''tr_fusion'' (refer to the train code) for prediction.

Citation

If you find this work useful in your research, please cite the paper:

@inproceedings{kim2021multimodalmem,
  title={Multi-Modality Associative Bridging Through Memory: Speech Sound Recollected From Face Video},
  author={Kim, Minsu and Hong, Joanna and Park, Se Jin and Ro, Yong Man},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={296--306},
  year={2021}
}