A Multi-task Learning Framework for Time-continuous Emotion Estimation from Crowd Annotations

We propose Multi-task learning (MTL) for time-continuous or dynamic emotion (valence and arousal) estimation in movie scenes. Since compiling annotated training data for dynamic emotion prediction is tedious, we employ crowdsourcing for the same. Even though the crowdworkers come from various demographics, we demonstrate that MTL can effectively discover (1) consistent patterns in their dynamic emotion perception, and (2) the low-level audio and video features that contribute to their valence, arousal (VA) elicitation. Finally, we show that MTL-based regression models, which simultaneously learn the relationship between low-level audio-visual features and high-level VA ratings from a collection of movie scenes, can predict VA ratings for time-contiguous snippets from each scene more effectively than scene-specific models.