The author proposes a contrastive learning method to address the scarcity of AU annotations by learning from unlabelled facial videos, achieving discriminative AU representations and person-independent detection.
