Comparative Evaluation of Modular Deep Learning Pipelines for Student Engagement Detection

Student engagement detection is crucial in e-learning environments where educators have limited direct visual feedback of learner attentiveness. This paper proposes a modular three-stage pipeline for automated engagement detection, consisting of face detection, emotion recognition, and head pose estimation. We implement and compare alternative models for the latter two stages (DeepFace vs. HSE for emotion, and Hopenet vs. ShuffleNet for pose) and demonstrate that the proposed pipeline using YOLOv8 + HSE + Hopenet achieves superior performance. In evaluations on the public DAiSEE dataset, our best pipeline reaches 90.25% binary classification accuracy with an F1-score of 94.85%, outperforming existing approaches to binary engagement detection on this benchmark. We further show via ablation studies that the choice of emotion recognition model has the greatest impact on accuracy, while the head pose estimator provides complementary improvements. By comparing several pipeline configurations, we provide insights into the modular pipeline's adaptability and performance.