GCC-FER dataset and CA-FER system tackle cultural bias in facial recognition
New 23,934-sample dataset reveals emotional expressions vary by culture.
A team of researchers has launched GCC-FER, a new hybrid multicultural video dataset for dynamic facial expression recognition (DFER) that challenges the assumption that emotions are universally expressed. The dataset contains 23,934 video samples spanning four cultural groups (African, Caucasian, East Asian, and South Asian) across seven basic expressions. It combines psychologically supervised in-house data collection for underrepresented populations with rigorous ethnicity filtering of existing sources, making it the first large-scale global cross-cultural DFER dataset.
Alongside the dataset, the team introduced the Culture-Aware FER (CA-FER) system, which uses behaviorally grounded cultural priors to adaptively recalibrate latent facial representations. Extensive experiments on GCC-FER and the existing DFEW benchmark showed that CA-FER consistently improves FER performance across multicultural settings. This work addresses a critical gap in affective computing, where most systems assume universal emotions despite known cultural differences in facial muscle activation patterns.
- GCC-FER dataset includes 23,934 video samples from four cultural groups.
- CA-FER system adaptively recalibrates facial representations to reduce cultural bias.
- Experiments on GCC-FER and DFEW show consistent performance improvements across cultures.
Why It Matters
Enables fair and accurate emotion AI across global populations, reducing bias in human-computer interaction systems.