ICIP 2019: Frame Attention Networks for Facial Expression Recognition in Videos pdf
Debin Meng, Xiaojiang Peng, Yu Qiao, etc.
Pillow == 6.2.0
numpy == 1.17.2
torch == 1.3.0
torchvision == 0.4.1
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.
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 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.
If you find this code useful in your research, please consider citing us:
@misc{1907.00193,
Author = {Debin Meng and Xiaojiang Peng and Kai Wang and Yu Qiao},
Title = {frame attention networks for facial expression recognition in videos},
Year = {2019},
Eprint = {arXiv:1907.00193},
url={https://github.com/Open-Debin/Emotion-FAN}
}