Facial Action Unit Detection
Overview
Facial Action Unit (AU) detection maps facial images or video to labels for visible muscle actions, often predicting several AUs at once and identifying subtle or brief movements. Methods use convolutional or transformer-based visual features, facial landmarks, temporal models, and graphs of AU co-occurrence; newer approaches also use language-based semantic priors, multimodal cues, and controlled synthetic data. Systems are commonly evaluated on BP4D and DISFA, with in-the-wild tests such as Aff-Wild2 and specialized protocols for micro-AUs. A central challenge is reliable generalization across people, datasets, and recording conditions, especially with scarce, imbalanced labels and evaluation results that can vary substantially across data splits.
Latest papers
23 papers1 paper in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.