Context-Aware Mixture of Domain Experts for Bodily Expression of Emotion in the Wild

The paper introduces Context-Aware Mixture of Domain Experts (CA-MoDE) for bodily emotion recognition in unconstrained scenes. It uses separate body, scene, and object experts, with the latter two producing emotion distributions that act as structured contextual priors. A task-specific max-endorsement gate selects the strongest contextual signal for each emotion dimension instead of averaging potentially conflicting distributions. CA-MoDE reports an Emotion Recognition Score of 0.3269 on the Body Language Database and surpasses existing temporal models while using only still images, suggesting that structured spatial context can partially proxy for behavioral dynamics.