A Tiny Vision-Based Model for Real-Time Student Attention Detection in Online Classes

Abstract

Online and blended classrooms widen access but remove the in-person cues instructors use to gauge attention. Prior work typically relies on heavy, cloud-bound or multimodal models that are hard to deploy on commodity laptops, treats attention as an unordered label without calibrated probabilities, and evaluates on subject-overlapping splits with limited robustness analysis. This creates a gap in Tiny, deployable, calibration-aware methods validated under realistic protocols. We address this gap with a TinyML, vision-only pipeline that estimates four attention levels: (Very Low, low, high, Very High ) from short webcam clips under strict on-device budgets. Each clip of T=30 frames at 224×224 is processed by a compact hybrid encoder: a CNN extracts per frame spatial features, a BiLSTM models temporal context, and a lightweight GRU refines dynamics; three parallel branches with staggered widths encourage feature diversity before fusion. We apply structured pruning of convolutional channels and recurrent units, post-training INT8 quantization, and temperature scaling for calibrated probabilities; models are exported as ONNX. On DAiSEE with subject-independent splits, the baseline attains 99.86% accuracy and 0.998 macro-F1, with strong ordinal agreement (QWK = 0.998, ordinal MAE = 0.03). The compressed model preserves reliability (macro-F1 = 0.995, QWK = 0.995), remains robust to low light, partial occlusion, and head yaw, and yields ∼4× smaller size and ∼2.3× CPU speedups. These results indicate a deployable, privacy-preserving approach to fine-grained, on-device attention analytics.