The proposed approach comprehensively models multi-scale facial dynamics and hierarchical spatio-temporal relationships among facial action units to achieve state-of-the-art performance in action unit occurrence recognition.
