Controllable student engagement recognition system : from modular vision pipelines embedding-aware, class-balance-tunable contrastive learning

Automated student engagement recognition supports adaptive online learning, but current sys tems face two challenges. First, visual cues of engagement (face presence, facial affect, and head pose) must be extracted reliably. Second, naturally collected data are severely imbal anced. Minority classes, such as Highly Disengaged and Highly Engaged, are rare, so high overall accuracy can hide poor recognition of these classes. This thesis investigates both problems. First, it introduces a three-stage modular pipeline (face detection, facial emotion recognition, and head-pose estimation) with four configurations tested on the DAiSEE benchmark for binary engagement classification. The best configuration achieves 90.25% accuracy and a binary F1-score of 0.9485. Because binary DAiSEE is 94.26% Engaged, these aggregate scores do not establish reliable Not Engaged detection. The study instead provides a controlled, interpretable comparison of cue-extraction modules. Second, the analysis shifts to the four-class ordinal CMOSE benchmark and representation level training. DUAL-POOL MOCORANK (DP-MR) replaces MocoRank’s single contrastive score pool with two pools: a class-balanced pool representing each engagement class equally and a natural-prior pool preserving the original class distribution. Embedding-aware post-hoc calibration (EMBED-LR) uses a held-out validation split to adjust the model without retraining. On CMOSE, calibration moves the model from 77.64% accuracy and 52.44%averageper-class accuracy to more balanced results: 71.50%/62.87% for general use, 65.93%/62.69% for flagging at-risk students, and 64.29%/64.37% at the most balanced setting. Bootstrap analysis shows that these changes are statistically reliable. In a matched single-model comparison, DP-MR improves on the published MocoRank with Center Loss by +1.88 percentage points in average per-class accuracy. Together, these contributions show that engagement recognition should not be reported as one fixed result. DP-MR can favor either overall accuracy or better detection of rare but impor tant disengagement classes. No setting is clearly better than the strongest published average per-class accuracy (60.94%). The main contribution is the ability to choose the best setting for the task, supported by an independent MocoRank reimplementation with documented replication details.