Student Engagement Recognition From Facial Landmark Trajectories Using Region-Aware Temporal Networks (RATE-Net)
IEEE Access · 2026 · 14 · 103790-103808
DOI: 10.1109/access.2026.3711637 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Rina Yulius und 2 weiteren Autor:innen. Die Arbeit nennt 75 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Indonesia Endowment Fund for Education (LPDP), Ministry of Finance of the Republic of Indonesia.
Abstract (Originalsprache)
Student engagement plays a crucial role in influencing attention, comprehension, and learning outcomes. However, tracking it automatically at scale remains difficult. Observing manually is impractical, and current vision-based systems depend on raw RGB data, which is highly costly and sensitive to variations. We propose a facial–landmark–based engagement recognition framework that models temporal dynamics using structured landmark trajectories. Instead of encoding the full-face holistically, we present a region-of-interest (ROI)–aware temporal modeling strategy that decomposes landmarks into three expressive regions: eyes, mouth, and eyebrows. Each region was processed by a dedicated Temporal Convolutional Network (TCN) encoder and subsequently fused for classification. A class-weighted loss was employed to address the label imbalance, and Macro-F1 was used as the primary evaluation metric. Under stratified 5-fold cross-validation, the proposed ROI-Aware Temporal Engagement Network (RATE-Net) achieves a Macro-F1 of $59.1\%\pm 0.3$ and accuracy of $66.0\%\pm 1.1$ , improving over a Global-TCN baseline (Macro-F $1~53.8\%\pm 2.0$ , accuracy $62.4\%\pm 1.2$ ). ROI ablation experiments showed that the eye region provided the strongest single-region signal (Macro-F $1~54.8\%\pm 1.0$ ), whereas multi-region fusion yielded the most stable and best overall performance. Cross-dataset evaluations further show that ROI-aware temporal modeling maintains a higher Macro-F1 than full-face representations under heterogeneous recording conditions. Finally, efficiency analysis showed that the ROI-fusion model reduced the parameter count and FLOPs by approximately 40–45% relative to the baseline while maintaining higher Macro-F1 performance. These results indicate that ROI-based temporal landmark modeling provides a compact and interpretable alternative to holistic landmark modeling, with consistent empirical improvements under the evaluated cross-validation setting.
Förderung
- Indonesia Endowment Fund for Education (LPDP), Ministry of Finance of the Republic of Indonesia
Zitieren
APA
Yulius, R., Sumpeno, S., & Purnomo, M. H. (2026). Student Engagement Recognition From Facial Landmark Trajectories Using Region-Aware Temporal Networks (RATE-Net). IEEE Access, 14, 103790-103808. https://doi.org/10.1109/access.2026.3711637
Vancouver
Yulius R, Sumpeno S, Purnomo MH. Student Engagement Recognition From Facial Landmark Trajectories Using Region-Aware Temporal Networks (RATE-Net). IEEE Access. 2026;14:103790-103808. doi:10.1109/access.2026.3711637
BibTeX
@article{Yulius2026,
author = {Yulius, Rina and Sumpeno, Surya and Purnomo, Mauridhi Hery},
title = {Student Engagement Recognition From Facial Landmark Trajectories Using Region-Aware Temporal Networks (RATE-Net)},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {103790--103808},
doi = {10.1109/access.2026.3711637}
}
EUVIMED