Shuffle Vision Transformer: Lightweight, Fast and Efficient Recognition of Driver’s Facial Expression

This is the official repository for the paper "Shuffle Vision Transformer: Lightweight, Fast and Efficient Recognition of Driver’s Facial Expression".

ShuffViT-DFER Architecture

Datasets

• KMU-FED dataset from https://cvpr.kmu.ac.kr/KMU-FED.html
• KDEF from https://kdef.se/download-2/index.html

Preprocessing

-For KMU-FED dataset: 'python preprocess_kmu.py' to save the data in .h5 format, then, "KMU.py" to split the data into 10 folds.
-For KDEF dataset: 'python preprocess_KDEF.py' to save the data in .h5 format, then, "KDEF.py" to split the data.

Train and Test model for all 10 fold

• KMU-FED dataset: python 10fold.py
• KDEF dataset: python combinedmodelkdef.py --model Ourmodel --bs 32 --lr 0.0001

plot confusion matrix

• python confmatrixkmu.py --model Ourmodel
• python confmatrixkdef.py --model Ourmodel

KMU-FED Accurary

We use 10-fold Cross validation in the experiment.

• Model： ShuffViT-DFER ; Average accuracy： 97.273%
  

KDEF Accurary

• Model： ShuffViT-DFER ; Accuracy： 92.441%
  

Confusion matrices