README

Introduction

This is the Pytorch Version for the reimplementation of Dual-stream Shallow Networks for Facial Micro-Expression Recognition. [Head to official code]

Files

• dataloader: Function for building custom dataset
• network: Network design
• preprocess: Function for processing the original dataset
• save_preprocess_weight: Save the preprocessed image into npz for repeating usage
• train: Train the model
• utils: Rename the fileanem in SAMM dataset

Package requirements

Install the required packages in this project by:

$ pip install -r requirements.txt

Usage

Note that for SAMM dataset, you should rename the data to fulfill the format in the code. To rename the dataset, see utils.py.

Preprocessing

The preprocessed images can be reused to cut down the training time. The usage is shown below:

usage: save_process_image.py [-h] --csv_path CSV_PATH --root ROOT --save_path
                             SAVE_PATH --catego CATEGO

optional arguments:
  -h, --help            show this help message and exit
  --csv_path CSV_PATH
  --root ROOT
  --catego CATEGO

For example:

$ python save_process_weight \
    --csv_path <path_to_csv_file> \
    --root <Path to images> \
    --catego <category>

Training

usage: train.py [-h] --path PATH --catego CATEGO --img_root IMG_ROOT [--num_classes NUM_CLASSES]
                [--combination_mode COMBINATION_MODE] [--image_mode IMAGE_MODE [IMAGE_MODE ...]]
                [--batch_size BATCH_SIZE] [--weight_save_path WEIGHT_SAVE_PATH] [--model MODEL]
                [--epochs EPOCHS] [--learning_rate LEARNING_RATE]

optional arguments:
  -h, --help            show this help message and exit
  --path PATH           Path for the csv file for training data
  --catego CATEGO       SAMM or CASME dataset
  --img_root IMG_ROOT   Root for the training image
  --num_classes NUM_CLASSES
                        Classes to be trained
  --combination_mode COMBINATION_MODE
                        Mode to be used in combination
  --image_mode IMAGE_MODE [IMAGE_MODE ...]
                        Image type to be used in training
  --batch_size BATCH_SIZE
                        Training batch size
  --weight_save_path WEIGHT_SAVE_PATH
                        Path for the saving weight
  --model MODEL         Model to used for training
  --epochs EPOCHS       Epochs for training the model
  --learning_rate LEARNING_RATE
                        Learning rate for training the model

For example:

$ python train.py \
    --path <path to csv file> \
    --img_root <place for your dataset> \
    --catego CASME \
    --image_mode F G \
    --weight_save_path CASME_DSSN_weight \
    --model DSSN

Especially, user can decide what mode to be used in training. DSSN needs two modes for training. On the other hand, SSSN only needs to choose one mode.

# For DSSN
... \
--image_mode F G \
...

# For SSSN
... \
--image_mode G

Pretrained AlexNet

The weight was obtained from YOUSIKI's GitHub repo.

Metrics

Model weights

Model	Paramters
SSSN	0.63M
DSSN	0.972M

Citation

@inproceedings{khor2019dual,
    title={Dual-stream Shallow Networks for Facial Micro-expression Recognition},
    author={Khor, Huai-Qian and See, John and Liong, Sze-Teng and Phan, Raphael CW and Lin, Weiyao},
    booktitle={2019 IEEE International Conference on Image Processing (ICIP)},
    pages={36--40},
    year={2019},
    organization={IEEE}
}