Project for Advanced Machine Learning 2018 (by Lennard Kiehl & Raphael Sayer)

Facial Expression Recognition 2013

Data sources

• original challenge (kaggle) (paper)
• dataset: FER-2013 (kaggle download)
• possible better annotations: FER+ (github) (paper)

Data description

• 35887 grayscale images (48x48x1) of faces (preprocessed)
• orignal labels include 7 classes/emotions (one per image): (0=Angry, 1=Disgust, 2=Fear, 3=Happy, 4=Sad, 5=Surprise, 6=Neutral)
• updated labels (FER+): 10 crowd-sourced labels per image
• split: train - 28709, public test/eval - 3589, test- 3589

Performance on original labels

Model	Top-1 Accuracy (%)
"null" model	60
ensembel of "null" models	65.5
original winner	71.162
approximate performance of vanilla CNN (found through browsing)	~64-68

Possible Architectures

• AlexNet (2012) (paper)
• VGG (2014) (paper)
• ResNet (2015) (paper)
• Highway Networks (2015) (paper)
• DenseNet (2016) (paper)
• SqueezeNet (2016) (paper)
• ConvNet-AIG (2017) (paper)
• MobileNetV2 (2018) (paper)