DFEW

Xingxun Jiang, Yuan Zong, Wenming Zheng, Chuangao Tang, Wanchuang Xia, Cheng Lu, Jiateng Liu "DFEW: A Large-Scale Database for Recognizing Dynamic Facial Expressions in the Wild". ACM MM'20

[Paper/中文版] [Download DFEW] [PPT] [poster] [video]

Requirements

• Python == 3.6.0
• PyTorch == 1.8.0
• Torchvision == 0.8.0

Training

• Step 1: download the single-labeled samples of DFEW dataset, and make sure it has the structure like the following:

/data/jiangxingxun/Data/DFEW/data_affine/single_label/
                                                      data/
                                                           00001/
                                                                 00001_00001.jpg
                                                                 ...
                                                                 00001_00144.jpg
                                                           16372/
                                                                 16372_00001.jpg
                                                                 ...
                                                                 16372_00039.jpg
                                                      label/
                                                            single_trainset_1.csv
                                                            ...
                                                            single_trainset_5.csv
                                                            single_testset_1.csv
                                                            ...
                                                            single_testset_5.csv
[Note]: 1:Happy 2:Sad 3:Neutral 4:Angry 5:Surprise 6:Disgust 7:Fear

• Step 2: run run.sh

Pretrained Models

• We provided the model's pretrained weights of each fold under 5-fold cross-validation protocol.

• To get the better metrics, we provided the models trained by extracted frames from videos, other than those acquired by the Time Interpolation Method with matlab code.

• We take two metrics to evaluate models, i.e., WAR and UAR. WAR is the weighted average recall, i.e., accuracy; and UAR is the unweighted average recall, i.e., the accuracy per class divided by the number of classes without considerations of instances per class.

[^_^]:- you can download model by wget command like wget https://aip.seu.edu.cn/download/pretrained_model/DFEW/scratch/<model_name>/, where <model_name> can be replaced with mc3_18, r3d_18, i3d ...

• you can download the pretrained weights from Baidu Disk with Access Code (8azt) or Google Driver. >

Citation

@inproceedings{jiang2020dfew,
  title={Dfew: A large-scale database for recognizing dynamic facial expressions in the wild},
  author={Jiang, Xingxun and Zong, Yuan and Zheng, Wenming and Tang, Chuangao and Xia, Wanchuang and Lu, Cheng and Liu, Jiateng},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={2881--2889},
  year={2020}
}