Engagement Detection in e-Learning Environments using Convolutional Neural Networks

Improving state-of-the-art in Detecting Student Engagement with Resnet and TCN Hybrid Network

“…Dewan et el. [25] and [26], modified the original four-level video engagement annotations in the DAiSEE dataset, and defined two and three-level engagement detection problems based on the labels of other emotional states in the DAiSEE dataset. They also changed the video engagement detection problem in the DAiSEE dataset to image engagement de-tection problem and performed their experiments on 1800 images extracted from videos in the DAiSEE dataset.…”

“…They achieved 90.89%, and 87.25% accuracy for two and three-level engagement detection, respectively. In [26], the extracted face regions are fed to different 2D CNN architectures to detect the level of engagement, and achieved 92.33% accuracy for two-level engagement detection. The authors in [25] and [26] have altered the original video engagement detection problem in the DAiSEE dataset [7] to image engagement detection problem and have not evaluated their methods on the video test set in the DAiSEE dataset (described in Section IV).…”

“…In [26], the extracted face regions are fed to different 2D CNN architectures to detect the level of engagement, and achieved 92.33% accuracy for two-level engagement detection. The authors in [25] and [26] have altered the original video engagement detection problem in the DAiSEE dataset [7] to image engagement detection problem and have not evaluated their methods on the video test set in the DAiSEE dataset (described in Section IV). Therefore, their reported accuracy would be hard to verify for generalization of results and their methods working with images cannot be compared to other works on the DAiSEE dataset that use videos.…”

The Asian Conference on Education 2020: Official Conference Proceedings

“…To train the model, we use the DAiSEE dataset with the feature extraction, in the same way, to extract the learners' face features while joining online learning. We use CNN because it is relatively simple and one of the deep learning methods broadly used in literature (Gudi et al, 2015;Li & Deng, 2020;Murshed et al, 2019;Nezami et al, 2017). Furthermore, we believe that simplicity and cost efficiency are the keys to a reliable implementation of engagement estimation in the actual online learning process.…”

“…Another limitation of this work is associated with the neural network we used for engagement estimation, where the result far from perfect. We acknowledge this limitation because using a typical CNN model that works with minimizing a loss function, which is computationally feasible but represents inaccurate prediction (Murshed et al, 2019). Some other features such as head pose, eye gaze, and distance between the monitor and the face can be further considered for input features for better accuracy.…”