ICIP 2019: Frame Attention Networks for Facial Expression Recognition in Videos pdf

Debin Meng, Xiaojiang Peng, Yu Qiao, etc.

If you are using pieces of the posted code, please cite the above paper.

We visualize the weights of attention module in the picture. The blue bars represent the self-attention weights and orange bars the final weights (the weights combine self-attention and relation-attention ).

Both weights can reflect the importance of frames. Comparing the blue and orange bars, the final weights of our FAN can assign higher weights to the more obvious face frames, while self-attention module could assign high weights on some obscure face frames. This explains why adding relation-attention boost performance.

Whether or not it is useful to you, could you please do not hesitate to citing the paper. thanks, thanks, thanks.:

@inproceedings{meng2019frame,
  title={frame attention networks for facial expression recognition in videos},
  author={Meng, Debin and Peng, Xiaojiang and Wang, Kai and Qiao, Yu},
  booktitle={2019 IEEE International Conference on Image Processing (ICIP)},
  pages={3866--3870},
  year={2019},
  organization={IEEE},
  url={https://github.com/Open-Debin/Emotion-FAN}
}

Pillow == 6.2.0

numpy == 1.17.2

torch == 1.3.0

torchvision == 0.4.1

We share two model of ResNet18, include a model pretrained in MS-Celeb-1M and another in FER+.Baidu or OneDrive

Training with self-attention

CUDA_VISIBLE_DEVICES=2 python Demo_AFEW_Attention.py --at_type 0

Training with self-attention and relation-attention

CUDA_VISIBLE_DEVICES=2 python Demo_AFEW_Attention.py --at_type 1

• --lr: initial learning rate
• --at_type: 0 is self-attention; 1 is relation-attention
• --epochs: number of total epochs to run
• --momentum: momentum
• --weight-decay: weight decay (default: 1e-4)
• -e: evaluate model on validation set
• etc.