Addressing the cold-start problem in facial expression recognition

Abstract

In our previous research [5] we proposed a CBR approach to infer the emotional state of the user through the analysis of a picture taken from the front facing camera of her mobile device. We demonstrated that different people express emotions with different gestures and got the best accuracy using a personal case base with self pictures of the same user. However, in the cold start situation, where pictures of the querying user are not available, the CBR system uses a generic case base (GCB) made of pictures of anonymous people. Although the performance using the GCB was acceptable on average there were several users with a very low accuracy. In this paper we compare our GCB to other reference picture catalogues and evaluate our CBR approach with state-of-the-art Facial Expression Recognition (FER) algorithms. Results point out that our approach is only suitable for GCB including semantically similar users. We use an ontology to group together users with similar demographic and physiological information: sex, age and ethnic group. We evaluate our CBR approach with small and specialized case bases where pictures are semantically similar to the target population and demonstrate that it efficiently increases the accuracy in the cold start situation and minimizes the noise in the case base.