QOT

QUATERNION ORTHOGONAL TRANSFORMER FOR FACIAL EXPRESSION RECOGNITION IN THE WILD

Requirements

• Python=3.8
• tensorflow=2.6.0
• PyTorch=1.10
• torchvision=0.11.0
• cudatoolkit=11.3
• matplotlib=3.5.3

Training & Evaluate

We evaluate QOT on RAF-DB, AffectNet and SFEW. We take RAF-DB as an example to introduce our method.

Traning

• Step 1: download RAF-DB datasets from official website, and put it into ./datasets
• Step 2: download pre-trained ResNet-50 from Google Drive, and put it into ./pretrianed
• Step 3: run main_Upload.py to train Orthogonal_CNN model.
• Step 4: replace the path with the pretrained model in Step 3 in main_generate_ortho.py to generate the numpy file of orthogonal features.
• Step 5: load orthogonal features generated in Step4 or directly download the pre-generated features from Google Drive, and run q-vit_RAFDB_Upload.py to training QOT module.

Evaluate

• Step 1: download RAF-DB datasets from official website, and put it into ./datasets
• Step 2: download the checkpoint from Google Drive, and put it into ./checkpoint_cnn
• Step 3: edit the evaluate_path with path in Step 2 and run main_Upload.py to evaluate Orthogonal_CNN model.
• Step 4: download the pre-generated orthogonal feature from Google Drive, and put it into ./orthogonal_npy
• Step 5: download the checkpoint from Google Drive, and put it into ./checkpoint_qvit
• Step 6: run the evaluate code in q-vit_RAFDB_Upload.py to evaluate QOT module.

Pre-trained Model

Orthogonal_CNN:Google Drive Orthogonal Features:Google Drive QOT: Google Drive