Building an Ecologically Valid Facial Expression Database – Behind the Scenes

Abstract

AbstractArtificial Intelligence (AI) algorithms, together with a general increased computational performance, allow nowadays exploring the use of Facial Expression Recognition (FER) as a method of recognizing human emotion through the use of neural networks. The interest in facial emotion and expression recognition in real-life situations is one of the current cutting-edge research challenges. In this context, the creation of an ecologically valid facial expression database is crucial. To this aim, a controlled experiment has been designed, in which thirty-five subjects aged 18–35 were asked to react spontaneously to a set of 48 validated images from two affective databases, IAPS and GAPED. According to the Self-Assessment Manikin, participants were asked to rate images on a 9-points visual scale on valence and arousal. Furthermore, they were asked to select one of the six Ekman’s basic emotions. During the experiment, an RGB-D camera was also used to record spontaneous facial expressions aroused in participants storing both the color and the depth frames to feed a Convolutional Neural Network (CNN) to perform FER. In every case, the prevalent emotion pointed out in the questionnaires matched with the expected emotion. CNN obtained a recognition rate of 75.02%, computed comparing the neural network results with the evaluations given by a human observer. These preliminary results have confirmed that this experimental setting is an effective starting point for building an ecologically valid database.KeywordsFacial expression recognitionEcologically-valid data3D facial databaseBasic emotionsAffective databaseHuman-robot interaction