- [pdf] [supp]
Domain Generalized RPPG Network: Disentangled Feature Learning with Domain Permutation and Domain Augmentation
Remote photoplethysmography (rPPG) offers a contactless method for monitoring physiological signals from facial videos. Existing learning-based methods, although work effectively on intra-dataset scenarios, degrade severely on cross-dataset testing. In this paper, we address the cross-dataset testing as a domain generalization problem and propose a novel DG-rPPGNet to learn a domain generalized rPPG estimator. To this end, we develop a feature disentangled learning framework to disentangle rPPG, identity, and domain features from input facial videos. Next, we propose a domain permutation strategy to further constrain the disentangled rPPG features to be invariant to different domains. Finally, we design a novel adversarial domain augmentation strategy to enlarge the domain sphere of DG-rPPGNet. Our experimental results show that DG-rPPGNet outperforms other rPPG estimation methods in many cross-domain settings on UBFC-rPPG, PURE, COHFACE, and VIPL-HR datasets.