Dynamic Facial Expression Recognition through Partial Label Learning and Federated Learning

Abstract

In this paper, we present a model development pipeline for dynamic Facial Expression Recognition (FER) aimed at quantifying learning in virtual classrooms. The proposed pipeline involves the use of partial labels for training dynamic FER models, followed by the use of a self-supervised federated learning approach in further enhancing the model's performance on new subjects, addressing both continual learning needs and privacy concerns. This work ultimately contributes to advancing learning quantification in virtual classrooms by integrating partial label training and federated learning strategies for dynamic FER.