Learning Facial Expression Recognition In-the-Wild from Synthetic Data Based on an Ensemble of Lightweight Neural Networks

This paper deals with one of the problems of recognizing the emotion from a photo gathered from in-the-wild settings, namely, facial expression recognition. We study various ensemble approaches that combine the lightweight MT-EmotiEffNet, attention-based transformer, and the graph-based model. The models have been implemented on an edge device (Jetson Nano) to be integrated into a demo application for video analytics to evaluate their performance in practical conditions. The experimental results demonstrated that the proposed approach outperformed competitive solutions using the Learning from Synthetic Data challenge datasets of the fourth Affective Behavior Analysis in-the-Wild competition.