Using fast.ai, an Image Classifier was trained with facial expression data from kaggle. The convnet architecture from timm was used to provide optimised training speed and model performance. Initial model accuracy was extremely poor and attempts at improving this by tuning the learning rate were semi-successfully, however, the issue was later realised to be caused by poorly labelled data. The model achieved a success rate of 70% although this poor accuracy did not affect its performance in the final application.
opencv and imutils were used to stream video from a computer web cam and pass the image tensor into the trained model for prediction. The prediction was then rendered over the video frame alongside a bounding box of the face and displayed in a window to the user.
- python 3.x
- Clone the repo:
git clone https://github.com/finnformic/video-facial-recognition.git- Change to working directory and setup python environment:
cd /foo/.../video-facial-recognition/
python3 -m venv venv
source venv/bin/activate- Install requirements:
pip3 install -r requirements.txt- Run application:
python3 main.py- Detects seven facial expressions (neutral, happy, sad, disgust, fear, surprise, angry)
- Frame skipping implemented to speed up video