On Automatically Assessing Children's Facial Expressions Quality: A Study, Database, and Protocol

While there exists a number of serious games geared towards helping children with ASD to produce facial expressions, most of them fail to provide a precise feedback to help children to adequately learn. In the scope of the JEMImE project, which aims at developing such serious game platform, we introduce throughout this paper a machine learning approach for discriminating between facial expressions and assessing the quality of the emotional display. In particular, we point out the limits in generalization capacities of models trained on adult subjects. To circumvent this issue in the design of our system, we gather a large database depicting children's facial expressions to train and validate the models. We describe our protocol to elicit facial expressions and obtain quality annotations, and empirically show that our models obtain high accuracies in both classification and quality assessment of children's facial expressions. Furthermore, we provide some insight on what the models learn and which features are the most useful to discriminate between the various facial expressions classes and qualities. This new model trained on the dedicated dataset has been integrated into a proof of concept of the serious game. Keywords: Facial Expression Recognition, Expression quality, Random Forests, Emotion, Children, Dataset