ACII'19 Apex-Time-Network (ATNet)

Code for the paper 'A Novel Apex-Time Network for Cross-Dataset Micro-Expression Recognition', presented at ACII'19.

Platforms and dependencies

Ubuntu 16.04 Python 2.7 CUDA8.0 CuDNN6.0+
Caffe：https://github.com/BVLC/caffe/

Prepare

• Download the database (option)
  CASMEII: http://fu.psych.ac.cn/CASME/casme2-en.php
  SAMM: http://www2.docm.mmu.ac.uk/STAFF/m.yap/dataset.php
  SMIC: https://www.oulu.fi/cmvs/node/41319

• if you wann't to download the original database, the Data fold contain all the needed data for this repositories.

  1. Add_python_layers contain a .py scrip that for image and point data load in Caffe.
  2. Apex_Cropped_images contains all the Apex images of three Database (namely: CASMEII SAMM SMIC). Apex_Cropped_images.txt contains image root and label in Apex_Cropped_images fold.
  3. OpticalFlowFeatureData.txt is the temporal features described in our paper.

ATN

This fold is our proposed network for Cross-Dataset Micro-Expression Recognition.

• CDE is the Composite Database Evaluation. For CASMEII_sub01, it means all samples (from the full consolidated database) are used for training except sub01 in CASMEII.

• HDE is the Holdout-Database Evaluation. For TEST_CASMEII, it means the model is trained on two datasets(SAMM SMIC) and tested on CASMEII.

• Notice: for each tries, you can use the get_samples_Train_Test_TXT.py to get the .txt list.

• How to run
  Take the CASMEII_sub01 fold for example, you just need to change your root in ../ACII19-Apex-Time-Network-master/, and then run: sh train_net.sh in the terminal.

MicroAttentionNet(Compared)

This fold is the compared method used in our paper.

Micro-Attention outperformed the method described in the paper "From Macro to Micro Expression Recognition: Deep Learning on Small Datasets Using Transfer Learning", the latter one won the first place in the Facial Micro-Expression Grand Challenge (MEGC2018) at FG 2018.

LBP-TOP(Compared)

This fold is the compared method used in our paper.

Features.zip contained all the LBP-TOP features.

HOOF(Compared)

This fold is the compared method used in our paper.

Features.zip contained all the HOOF features.
The optical method for HOOF is refer to https://www.researchgate.net/publication/320373402_Dual_Temporal_Scale_Convolutional_Neural_Network_for_Micro-Expression_Recognition

Citation

If it is helpful, please cite:

Peng, Min, Chongyang Wang, Tao Bi, Yu Shi, Xiangdong Zhou, and Tong Chen. "A novel apex-time network for cross-dataset micro-expression recognition", In 2019 8th International Conference on Affective Computing and Intelligent Interaction (ACII), pp. 1-6. IEEE, 2019.