Abstract
Emotion recognition models are widely used in affective computing and learning analytics. However, most models are trained on web-scale or controlled datasets that differ substantially from real classroom environments. Classrooms introduce variations in camera angle, lighting, image resolution, and interaction structure, which may affect both prediction reliability and fairness across student populations. This paper evaluates emotion recognition models in classroom learning environments. We analyze three stateof-the-art models (EmoNet, HSEmotion, and iMotions) using data collected from 69 students across three learning environments: game-based, embodied, and collaborative activities. The evaluation examines model sensitivity to visual perturbations and demographic variation across Fitzpatrick skin tone groups. Results show that prediction errors increase under changes in camera angle, lighting, and resolution, and vary systematically across skin tone groups.
We further observe that embodied learning environments exhibit higher sensitivity to visual perturbations compared to game-based and collaborative settings. Lightweight mitigation strategies using data augmentation and partial fine-tuning reduce both prediction error and fairness gaps. These findings highlight challenges in deploying emotion recognition models in classroom learning environments and motivate evaluation practices that account for real-world classroom conditions.
