Crowd Sentiment Analysis Project

Across recent scholarship, facial expression research has converged on a small set of benchmark datasets and architectural motifs that can be directly repurposed for analysing classroom video.[:cite[1]{ln=5}] ==FER2013 and CK+ remain the de facto evaluation standards, while Spontaneous ME corpora like CASME and SMIC supply the fine grained cues essential for micro expression work.[:cite[3]{ln=6}][:cite[2]{ln=3}]== These resources pair with deep learning backbones—often fused with attention blocks or detection first pipelines such as YOLOxx—to move FER from curated images to in the wild classrooms.[:cite[5]{ln=1}][:cite[4]{ln=6}] Standard datasets drive method development • FER2013 delivers 35 887 labeled, low resolution images drawn from internet videos.[:cite[5]{ln=6}] • Seven emotions (anger, disgust, fear, happiness, sadness, surprise, neutral).[:cite[6]{ln=6}] • Pose and lighting diversity create “hard negatives” that test robustness.[:cite[5]{ln=3}] • CK+ complements FER2013 with 327 high resolution sequences showing the seven basic emotions performed by 123 subjects in studio settings.[:cite[6]{ln=1}][:cite[6]{ln=3}] • Spontaneous micro expression corpora (CASME, CASME II, SMIC, SAMM) contain 195–203 ultra fast clips at 60–200 fps, capturing concealed emotions in naturalistic elicitation protocols.[:cite[7]{ln=2}][:cite[2]{ln=4}] • Table 1 contrasts scope and scale. Dataset Size Capture Emotions Focus Citation FER2013 35.9 k images In the wild 7 Macro FER [:cite[5]{ln=6}] CK+ 327 sequences Studio 7 High quality [:cite[6]{ln=3}] CASME 195 clips Spontaneous (60 fps) 7 MEs only [:cite[7]{ln=2}] SMIC 77 clips Spontaneous (100 fps) 3 MEs only [:cite[2]{ln=3}] Table 1 – Key FER & ME datasets Architecture families moving into the classroom • Single stage detectors (YOLO versions 5–12) add express recognition to real time, low latency pipelines.[:cite[5]{ln=2}][:cite[4]{ln=6}] • YOLOv11n/12n trained on FER2013/KDEF show mAP of 65.2 %–95.6 %, trading precision for speed.[:cite[9]{ln=4}][:cite[8]{ln=2}] • YOLOv8 fine tuned for MEs attains 70 % on in lab data and sustaining 80 % on test splits.[:cite[10]{ln=2}][:cite[11]{ln=9}] • Backbone+attention hybrids (ResNet50 + CBAM + TCN) reach 91.7–97.1 % on FER2013/CK+, modelling temporal cues in live streams.[:cite[13]{ln=2}][:cite[12]{ln=2}] • Transformer hybrids such as MobileNetV3 ViT with token downsampling cut computational cost by 40 % while boosting accuracy 5.5 % over CNN alone baselines on FER2013.[:cite[14]{ln=6}][:cite[15]{ln=1}] • Multi head attention networks like DAN distribute focus over eyebrows, eyes, mouth, and chin, improving class separability on RAF DB and AffectNet.[:cite[17]{ln=5}][:cite[16]{ln=3}] Feature and learning strategies • Classical or deep features supplied to lightweight classifiers (LBP + KNN and LSTM) still reach near perfect F1 on CK+.[:cite[18]{ln=3}][:cite[18]{ln=6}] • Deep attention layers enable one shot end to end training, removing the detection–classification silo that was common two years ago.[:cite[19]{ln=2}][:cite[20]{ln=4}] • Temporal convolutions, self attention, and recurrent cells afford the granularity demanded by confusion states such as boredom or surprise within lectures.[:cite[13]{ln=6}][:cite[21]{ln=4}] Dataset limitations and mitigation • Scarcity of labeled classroom footage forces researchers to rely on lab or online datasets; performance drops when models transfer directly to real classes.[:cite[22]{ln=7}][:cite[23]{ln=3}] • Synthetic augmentation (morphs, noise, occlusion) and cross domain training (FER2013 + lab classroom data) raise Average Precision by 20 % on proprietary classroom tests.[:cite[23]{ln=2}] • Mask occlusion studies show disgust, happiness, and sadness most impaired, underscoring need for region completion or sensor fusion when masks are present.[:cite[24]{ln=6}] Open challenges • Spatiotemporal spotting: locating micro expressions within continuous lecture videos remains unsolved due to annotation costs.[:cite[25]{ln=1}] • Model bias: demographic and lighting imbalance in FER2013 may skew detections; fair representation is critical in diverse classrooms.[:cite[5]{ln=3}] • Real time scaling: edge deployment is still limited by model size; token downsampling and knowledge distillation are promising shortcuts.[:cite[15]{ln=3}][:cite[15]{ln=5}] • Ethical use: student consent, data privacy, and interpretation guidelines must accompany any live emotion monitoring system, an area rarely addressed in current work.[:cite[26]{ln=8}] Diagram: Typical pipeline for classroom FER Synthesis mind map Takeaways FER in classrooms leverages public macro datasets and niche micro expression corpora alike; modern models blend detection, attention, and temporal modules to edge towards teacher ready prototypes.[:cite[3]{ln=6}][:cite[13]{ln=2}] Real world deployment hinges on richer—and more ethically collected—classroom data plus continued strides in lightweight, attention integrated backbones.[:cite[23...