Implementation of DeepEmotion paper

• Network to detect facial emotion expression using CNN and Spatial Transformers.

Dataset

• Download and prepare FER 2013 dataset
  • 48x48 pixel grayscale images of faces
  • emotion shown in the facial expression in to one of seven categories (0 = Angry, 1 = Disgust, 2 = Fear, 3 = Happy, 4 = Sad, 5 = Surprise, 6 = Neutral)
  • The file icml_face_data.csv contains the train(28709 images), validation(3589 images) and test data(3589 images).

  kaggle competitions download -c challenges-in-representation-learning-facial-expression-recognition-challenge
  unzip challenges-in-representation-learning-facial-expression-recognition-challenge.zip

Training

• using torch - 2.1.0

  python train.py --log_dir logs -e 100 -b 256
  python test.py --model deep_emotion-256-0.005-300.pt

• Test accuracy of 51% after training for 300 epochs. But the paper claims 70.1% accuracy.
• Some information is missing in the paper like the batch size. Going by the Spatial transformer configuration, the input to FC layer would be 1000 instead of 90.

Live prediction using webcam

• opencv configuration file to be downloaded from here

  python cam.py --model deep_emotion-256-0.005-300.pt