Efficient Algorithms for Video-Based Engagement Prediction for a MOOC Course
Polina Demochkina, Andrey Savchenko · 2022 International Russian Automation Conference (RusAutoCon) · 2022
In this paper, we address the engagement prediction task using a two-stage approach. First, two types of features are extracted from each frame on the input video using the OpenFace toolkit and an EfficientNet-based model that was pre-trained for the task of age, gender, and identity prediction. Each set of obtained feature vectors is then separately aggregated into a single video descriptor using statistical functions and a prediction is made using a linear support vector machine. Finally, the scores are fused, and the engagement intensity is predicted. The described ensemble model achieves MSE of 0.0534 on the test set of the EngageWild dataset from the EmotiW 2020 challenge.