Informative Feature Disentanglement for Unsupervised Domain Adaptation

Unsupervised Domain Adaptation (UDA) aims at learning a classifier for an unlabeled target domain by transferring knowledge from labeled source with related but different distribution. The strategy of aligning the two domains in latent feature space via metric discrepancy or adversarial has achieved considerable progress. However, these existing approaches mainly focus on adapting entire image and ignore bottleneck that occurs when forced adaptation uninformative domain-specific variations undermines effectiveness learned features. To address this problem, we propose novel component called Informative Feature Disentanglement (IFD), which is equipped network model, respectively. Accordingly, new architectures, named IFDAN IFDMN, enable informative refinement before adaptation. proposed IFD designed to disentangle features variations, are produced Variational Autoencoder (VAE) lateral connections encoder decoder. We cooperatively apply conduct supervised disentanglement unsupervised domain. In way, disentangled details Extensive experimental results three gold-standard datasets, e.g., Office31, Office-Home VisDA-C, demonstrate IFDMN models UDA.