README

Introduction

This repo is the source code for the paper: "Mimicking the Annotation Process for Recognizing the Micro Expressions".

Installation

• Set up the environment

  # Requires Python >= 3.8
  $ pip install -r requirements.txt

• Download landmarks file from dlib

  $ wget https://github.com/davisking/dlib-models/raw/master/shape_predictor_68_face_landmarks.dat.bz2 -P weight
  $ bzip2 -d weight/shape_predictor_68_face_landmarks.dat.bz2 

• Download the pretrained weights

  $ bash download.sh

Preprocessing

Face cropping

# Crop the CASME II dataset
#   We use OpenFace to crop the CASME II dataset
#   Refer to: https://github.com/TadasBaltrusaitis/OpenFace 

# Crop the SAMM dataset
$ python utils/samm_crop.py \
  --ori_path <original SAMM root> \
  --new_path <new path to save the crop image>

Magnified the image frame

We use Learning-based Motion Magnification to amplify our image frame. Follow the setup and install the requirements (recommend opening a new environment and not mix up with the previous one). Put the files in magnifed to deep_motion_mag folder.

# Set up deep_motion_mag
$ <clone the deep_motion_mag and setup the environment>

# Create magnified frame for CASME II and SAMM
$ python get_casme_img.py \
  --img_root <img root for CASME II dataset> \
  --csv_file ../csv_file/CASME.csv

$ python get_samm_img.py \
  --img_root <img root for SAMM dataset> \
  --csv_file ../csv_file/SAMM.csv

Create the On-A and Off-A frames

$ python utils/crop_face_optical.py \
  --file_name csv_file/CASME.csv \
  --img_root <img root of csv file> \
  --catego casme

$ python utils/crop_face_optical.py \
  --file_name csv_file/SAMM.csv \
  --img_root <img root of csv file> \
  --catego samm  

Training

The training file is put in the train_bash folder. Users should change the image root in the bash file first.

# Training five-class and three-class CASME II and SAMM
$ bash train_bash/five_casme.sh
$ bash train_bash/five_samm.sh
$ bash train_bash/three_casme.sh
$ bash train_bash/three_samm.sh

Testing

To test our training weights, users should download the files in the Installation Section.

# Test for the five-class CASME II
$ python test.py \
  --img_root <img root of CASME II> \
  --csv_file csv_file/sep_five_casme.csv \
  --weight_root weight/five_casme_best \
  --catego casme \
  --num_classes 5

# Test for the five-class SAMM
$ python test.py \
  --img_root <img root of SAMM> \
  --csv_file csv_file/sep_five_samm.csv \
  --weight_root weight/five_samm_best \
  --catego samm \
  --num_classes 5

# Test for the three-class CASME II
$ python test.py \
  --img_root <img root of CASME II> \
  --csv_file csv_file/sep_three_casme.csv \
  --weight_root weight/three_casme_best \
  --catego casme \
  --num_classes 3

# Test for the three-class SAMM
$ python test.py \
  --img_root <img root of SAMM> \
  --csv_file csv_file/sep_three_samm.csv \
  --weight_root weight/three_samm_best \
  --catego samm \
  --num_classes 3

Our results are shown in the below table:

Visualization

• Distribution plot

  # For CASME II distribution plot
  $ python utils/plot_dist.py \
  --img_root <img root of CASME II> \
  --csv_file csv_file/sep_five_casme.csv \
  --catego casme \
  --weight_path weight/five_casme_best/model_best_1.pt \
  --save_fig <place to generate the plot>
  
  # For SAMM distribution plot
  $ python utils/plot_dist.py \
  --img_root <img root of SAMM> \
  --csv_file csv_file/sep_five_samm.csv \
  --catego samm \
  --weight_path weight/five_samm_best/model_best_6.pt \
  --save_fig <place to generate the plot>

• AU-CAM

  # For CASME II AU-CAM
  $ python utils/au_cam.py \
  --img_root <img root of CASME II> \
  --csv_file csv_file/sep_five_casme.csv \
  --catego casme \
  --weight_path weight/five_casme_best/model_best_1.pt \
  --num_classes 5
  --save_path <place to save the AU-CAM>
  
  # For SAMM AU-CAM
  $ python utils/au_cam.py \
  --img_root <img root of SAMM> \
  --csv_file csv_file/sep_five_samm.csv \
  --catego samm \
  --weight_path weight/five_samm_best/model_best_6.pt \
  --num_classes 5
  --save_path <place to save the AU-CAM>

Distribution Plot

CASME II

SAMM

Citation

@inproceedings{ruan2022mimicking,
  title={Mimicking the Annotation Process for Recognizing the Micro Expressions},
  author={Ruan, Bo-Kai and Lo, Ling and Shuai, Hong-Han and Cheng, Wen-Huang},
  booktitle={ACM International Conference on Multimedia},
  year={2022}
}