In December 2017 I embarked on a journey to critically review the literature on facial expression recognition (FER) and convolutional neural networks (CNNs). After two months of determined reviewing I had read all of the literature I could find. I read published articles and pre-prints and did not limit myself to any journal database. The next three months were dedicated to writing the literature review and conveying my thoughts in a critical manner.
I presented some of my findings at the Deep Learning Indaba X Western Cape 2018. The link to my talk is available on YouTube here and my slide deck is available here. I first introduce the audience to facial expression recognition, mention applications and discuss some of the primary steps to consider when implementing a CNN for FER. I did not discuss additional aspects such as network and hyper-parameter optimisation. These are discussed in great detail in my article which is currently under review.
I then gave the talk again with a slight variation at AIMS South Africa. In this talk I presented the questions which I used to guide my critical analysis. Writing in a critical manner, and not simply summarising text is challenging. Authors often tend to summarise existing literature. This will be the topic of one of my future posts. The slides that I presented at the talk at AIMS are available here
The abstract to my talk is as follows:
We use facial expressions to convey happiness, sadness and many other emotions. Humans are good at understanding these universal expressions, however, this is non-trivial for a machine. There has not yet been significant effort at challenging the existing literature and to critically analyse what has been done and what should be improved upon in future research. In this talk we will discover what has been done in this field with respect to each primary aspect of pre-processing, building, training and evaluating convolutional neural networks (CNNs) for facial expression recognition (FER). The talk is targeted to those who wish to learn about the theoretical aspects to be considered when implementing a CNN from scratch for FER. No background knowledge in FER is required.