Zengqun Zhao, Qingshan Liu, Feng Zhou. "Robust Lightweight Facial Expression Recognition Network with Label Distribution Training". AAAI'21

• Python >= 3.6
• PyTorch >= 1.2
• torchvision >= 0.4.0

• Step 1: download basic emotions dataset of RAF-DB, and make sure it has the structure like the following:

./RAF-DB/
         train/
               0/
                 train_09748.jpg
                 ...
                 train_12271.jpg
               1/
               ...
               6/
         test/
              0/
              ...
              6/

[Note] 0: Neutral; 1: Happiness; 2: Sadness; 3: Surprise; 4: Fear; 5: Disgust; 6: Anger

• Step 2: download pre-trained model from Google Drive, and put it into ./checkpoint.
• Step 3: change the --data in run.sh to your path
• Step 4: run sh run.sh

• Add test and visualization code. (May. 5, 2023 Update)

• Sept. 16, 2021 Update
  We provide the pre-trained ResNet-18 and ResNet-50 on MS-Celeb-1M (classes number is 12666) for your research.
  The Google Driver for ResNet-18 model. The Google Driver for ResNet-50 model.
  The pre-trained ResNet-50 model can be also used for LDG.
• Nov. 6, 2021 Update
  The fine-tuned LDG models on CAER-S, AffectNet-7, and AffectNet-8 can be downloaded here, here, and here, respectively.
• Nov. 12, 2021 Update
  The trained EfficientFace model on RAF-DB, CAER-S, AffectNet-7, and AffectNet-8 can be downloaded here, here, here, and here, respectively. As demonstrated in the paper, the testing accuracy is 88.36%, 85.87%, 63.70%, and 59.89%, respectively.

@inproceedings{zhao2021robust,
  title={Robust Lightweight Facial Expression Recognition Network with Label Distribution Training},
  author={Zhao, Zengqun and Liu, Qingshan and Zhou, Feng},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={35},
  number={4},
  pages={3510--3519},
  year={2021}
}

The samples' number of the CAER-S dataset employed in our work should be: all (69,982 samples), training set (48,995 samples), and test set (20,987 samples). We apologize for the typos in our paper.