A novel unsupervised domain adaptation approach for action recognition from videos, inspired by recent literature on contrastive learning. It comprises a novel two-headed deep architecture that simultaneously adopts cross-entropy and contrastive losses from different network branches to robustly learn a target classifier.