Face Emotion Recognition in SVM

A simple facial emotion recognition classifier implemented by SVM

prepare work

Create an environment for the experiment using the following command

conda create -n SVM python=3.8
conda activate SVM
pip install -r requirements.txt

Download the corresponding data set from Kaggle

try to use

The structure of our model is as follows

SVM
	|
	| - log (our experiment result)
	| - data (you should get this directory from kaggle)
	| - model (some models we use to extract feature)
		| - hog.py
		| - resnet.py
		| - ...
	| - main.py
	| - ...

We provide detailed parameter descriptions in main.py

train and eval

python -u main.py --kernel rbf --method {your_method} --gpu_id {your_gpu_id} --C 5 --gamma 0.02 > res.log

our experiment result

kernel	feature method	other description	score
sigmoid	/	/	24.97
rbf	/	/	30.90
linear	/	/	not converged
linear	hog	patch(8,2)	45.31
rbf	hog	patch(8,2)	45.90
rbf	hog	patch(4,4)	51.46
rbf	hog+pca	patch(4,4)	51.52
rbf	align	dim: 11664->136	44.10
rbf	hog	C=5, gamma=0.02	57.65
rbf	hog+aug	C=5, gamma=0.02	59.07
rbf	resnet18+aug	C=5, gamma=0.02	68.04

for more details, please see ./log directory.