Introduction. Recognizing human facial expression from non-frontal, high-angle surveillance viewpoints has broad real-world applications, such as student mental health monitoring on campus and detection of abnormal psychological risks in public space. Traditional emotion recognition datasets, such as CK+, RAF-DB, AffectNet, are predominantly collected from a frontal perspective. This limits their ecological validity due to the failure in capturing the feature distortion and occlusion inherent in overhead views, and consequently reduces model generalizability in surveillance scenarios.
Purpose. This study aims to bridge this gap by constructing a high-angle facial expression dataset reflecting realistic observation conditions. This dataset is based on Ekman's basic emotion theory, which posits universally recognizable emotions, including anger, disgust, fear, happiness, sadness, surprise, and neutral.
Method. Professional actors were recruited and guided with scenario-based instructions to evoke naturalistic emotional expressions, producing authentic facial responses that can be readily captured. Two high-angle surveillance cameras with 1920×1080 resolution simultaneously recorded the actors' facial expressions. All facial expressions in the dataset were annotated by rigorously trained psychological experts, ensuring reliable labeling.
Results. Intra-rater consistency, assessed through repeated classifications, ranged from 0.819 to 0.962 (M = 0.89) indicating that each expert's coding was highly reliable. Images achieving a consensus code from at least six experts were retained, yielding 598 samples. In the final dataset, intra-rater consistency averaged 0.94 across images, with 75.9% of images showing consistency ≥ 0.90 and 52.5% reaching perfect agreement (1.0). Non-consensus samples were excluded to ensure high labeling reliability.
Conclusions. This validated High-Angle Facial Expression Dataset (HA-FED) enables investigation of how facial expressions manifest under atypical viewpoints, informing both theoretical understanding and practical applications. It offers a benchmark for developing robust affective computing models. HA-FED extends the field beyond conventional frontal facial expression datasets, supporting the translation of basic emotion theory into more applied, real-world contexts.