Class Overwhelms: Mutual Conditional Blended-Target Domain Adaptation

Current methods of blended targets domain adaptation (BTDA) usually infer orconsider domain label information but underemphasize hybrid categorical featurestructures of targets, which yields limited performance, especially under thelabel distribution shift. We demonstrate that domain labels are not directlynecessary for BTDA if categorical distributions of various domains aresufficiently aligned even facing the imbalance of domains and the labeldistribution shift of classes. However, we observe that the cluster assumptionin BTDA does not comprehensively hold. The hybrid categorical feature spacehinders the modeling of categorical distributions and the generation ofreliable pseudo labels for categorical alignment. To address these, we proposea categorical domain discriminator guided by uncertainty to explicitly modeland directly align categorical distributions $P(Z|Y)$. Simultaneously, weutilize the low-level features to augment the single source features withdiverse target styles to rectify the biased classifier $P(Y|Z)$ among diversetargets. Such a mutual conditional alignment of $P(Z|Y)$ and $P(Y|Z)$ forms amutual reinforced mechanism. Our approach outperforms the state-of-the-art inBTDA even compared with methods utilizing domain labels, especially under thelabel distribution shift, and in single target DA on DomainNet. Source codesare available at\url{https://github.com/Pengchengpcx/Class-overwhelms-Mutual-Conditional-Blended-Target-Domain-Adaptation}.