This is the official PyTorch implementation of the paper "Bridging the Gaps: Utilizing Unlabeled Face Recognition Datasets to Boost Semi-Supervised Facial Expression Recognition".
| Dataset | homepage | train size | test size |
|---|---|---|---|
| RAF-DB | http://www.whdeng.cn/raf/model1.html | 12271 | 3068 |
| FERPlus | https://github.com/microsoft/FERPlus | 287401 | 3500 |
| AffectNet-7 | http://mohammadmahoor.com/affectnet/ | 291651 | 4000 |
| AffectNet-8 | http://mohammadmahoor.com/affectnet/ | 28709 | 7178 |
Pretrain model weights is available checkpoints
| Method | Dataset | Accuracy | weights |
|---|---|---|---|
| SSFER-5% | RAF-DB | 80.96 | checkpoints |
| SSFER-10% | RAF-DB | 85.07 | checkpoints |
| SSFER-25% | RAF-DB | 88.23 | checkpoints |
| SSFER-5% | FERPlus | 81.77 | checkpoints |
| SSFER-10% | FERPlus | 83.95 | checkpoints |
| SSFER-25% | FERPlus | 85.82 | checkpoints |
| SSFER-1% | AffectNet-7 | 59.08 | checkpoints |
| SSFER-5% | AffectNet-7 | 64.02 | checkpoints |
| SSFER-10% | AffectNet-7 | 65.37 | checkpoints |
| SSFER-1% | AffectNet-8 | 53.85 | checkpoints |
| SSFER-5% | AffectNet-8 | 60.17 | checkpoints |
| SSFER-10% | AffectNet-8 | 61.65 | checkpoints |
- Clone the repository:
git clone https://github.com/{USERNAME}/{REPO_NAME}.git
cd SSFER- Create a conda environment and install the required dependencies:
conda create -n {ENV_NAME} python=3.10
conda activate {ENV_NAME}
pip install -r requirements.txtFinally, these files should be organized as follows:
├── RAF-DB
│ ├── train
│ └── test
├── FERPlus
│ ├── train
│ └── test
├── AffectNet-7
│ ├── train
│ └── test
├── AffectNet-8
│ ├── train
│ └── test
├── outputs_dir
│ ├── SSFER_FERPlus_0.05_model.pth
│ ├── SSFER_FERPlus_0.10_model.pth
│ ├── SSFER_FERPlus_0.25_model.pth
│ ├── SSFER_RAFDB_0.05_model.pth
│ ├── ......
│ ├── SSFER_AffcetNet8_0.01_model.pth
│ ├── SSFER_AffcetNet8_0.05_model.pth
│ ├── SSFER_AffcetNet8_0.10_model.pth
│ ├── 125w_base_warmup50_batch256_600.pth
│ ├── yolov7s-face.pt
├── utils
│ ├── ......
├── MAE_model.py
├── test.ipynb
The yolo-face repository used in the article is here.
Bibtex Citation