Abstract
In this paper, we address the Online Unsupervised Domain Adaptation (OUDA)problem, where the target data are unlabelled and arriving sequentially. Thetraditional methods on the OUDA problem mainly focus on transforming eacharriving target data to the source domain, and they do not sufficientlyconsider the temporal coherency and accumulative statistics among the arrivingtarget data. We propose a multi-step framework for the OUDA problem, whichinstitutes a novel method to compute the mean-target subspace inspired by thegeometrical interpretation on the Euclidean space. This mean-target subspacecontains accumulative temporal information among the arrived target data.Moreover, the transformation matrix computed from the mean-target subspace isapplied to the next target data as a preprocessing step, aligning the targetdata closer to the source domain. Experiments on four datasets demonstrated thecontribution of each step in our proposed multi-step OUDA framework and itsperformance over previous approaches.