Face-Sentiment

A workflow that combines face detection and sentiment classification of 7 concepts: Anger, Disgust, Fear, Happiness, Neutral, Sadness-Contempt, Surprise.

Face-Sentiment Visual Classifier Model:

• Training Data

  The training data is sourced from FER images (https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data) using the FER+ labels. In addition, Clarifai had relabelled some images by hand to accommodate to the current taxonomy.

• Performance Metrics

  Using FER out-of-sample eval set:

     Average ROC AUC: 0.9538
     
     Average f1: 0.7078
     
     Average precision: 0.7476
     
     Average recall: 0.682
  

- Model performs best on a cropped face

    The following concepts have an increased chance for confusion because of their visual similarities:

        fear and surprise
        
        neutral and sadness-contempt

- Other known limitations

    Silly faces, where the tongue is out; occluded faces
    
    Currently the model does not support a face showing pain/pleasure/arousal