Children's emotions dataset: Facial images as action units and valence scores

Abstract

The paper presents a dataset of emotions from children between 10 and 12 years old. This dataset was obtained from videos that are represented in time series of facial Action Units (AUs), and their corresponding valences were scored by professionals. The AUs are extracted from the videos using the Deepface library, and the valence series are obtained from expert observers who rate each video on a range from -1 to 1, covering the spectrum of negative to positive emotions. The dataset was evaluated by a total of 20 professional experts, comprising psychologists and psychology practitioners, with each video receiving an average of 10 reviews. The analysis encompassed a total of 57 videos, representing 22 students, culminating in the acquisition of a comprehensive set comprising 50 temporal series of action units and their associated weighted valence scores. This dataset is useful for training machine learning models in the process of identifying emotions to determine possible patterns of behaviour in classrooms. These patterns may reveal problematic academic attitudes or situations, or, conversely, the early identification of positive emotions that can empower leading students. In addition, it can assist education professionals in undertaking self-evaluations of their formative processes, with a focus on the emotions or attention exhibited by their students within the classroom environment during lessons.