• Face Dectection is accomplished by mediapipe developed by Google.

• Facial Expression Recognition is trained using DCNN (Deep Convolutional Neural Network) with FER+ dataset which is held by Microsoft.

(Thanks to @zc of the image)

• Language: python3.6

• Dependencies:

  • pytorch
  • opencv-python
  • mediapipe (modified)
  • CUDA10.1 (optional)
  • ...

• you may install all the dependencies via command python -m pip install -r requirements.txt

• Usage:

  • Run from prebuilt exe: see release-windows-v0.1 (Recommended)
  • Run from source:
    • Replace drawing_utils.py in mediapipe with src/drawing_utils.py in which I slightly modified.
    • Contact me by Email to get the trained model.

    python camdemo.py --camera 0

• Performance:

  • Absolutely REALTIME! The model could achieve above the average 60 FPS on a plain PC. If possible, try using a GPU to gain better performance!
  • My poor computer: Intel i7-7700K CPU (4.2GHz) with NVIDIA Quadro P2000 (5G memory)

• Model Structure: the model is quite simple though. It uses ResNet50 as backbone for feature extraction after which it is stacked with two fully connected layer. The output is a 10-size digits vector corresponding to 10 emotion classes.

• Accuracy: the model achieves 79.8% accuracy evaluated by FER+ valid subset after 14 epochs of training using softCE loss.

  epoch	KLdiv	softCE	weightedSoftCE
  0	0.005	0.005	0.005
  1	0.55	0.598	0.56
  2	0.58	0.652	0.668
  3	*	0.695	0.697
  4	*	0.726	0.71
  5	*	0.753	0.68
  6	*	0.76	0.665
  ...	...	...	...
  14	*	0.798	0.742

• Loss function:

  • Rather than original FER, each image in FER+ has been labeled by 10 crowd-sourced taggers but the default implementation of cross-entropy in pytorch uses just one hard label to compute the loss which abandons the information of 10 soft labels. So I implemented the soft cross-entropy to train the model fitting the probability distribution of emotion class which got pretty good results.

  • One more reason to use softCE loss is that for emotion classification, some human emotions cannot be distinguished well such as happiness and surprise.

  • As FER+ is a very imbalanced dataset (see image below) so I've tried use weightedSoftCE( like the idea of focal loss) but no good which I don't quite get it yet. If you happen to know why, tell me! Also when using weightedSoftCE during training, I found the loss rising upside and down a lot, which means it's not that numerical stable.

• Online education for children, which could be used to identify whether children listen carefully; For on-site meeting or school classroom, to judge the quality of the speech.

• On-site Human–Machine Interaction.