A Unified Framework for Covariance Adaptation with Multiple Source Domains

This paper addresses the problem of unsupervised domain adaptation in a setup where a single source is not sufficient for training the model. In this situation, a hybrid, multi-source driven training dataset is used. This calls for the need of an effective method to align the geometrically quasi-related source domains which would help prepare a better ground for aligning the unlabeled target dataset. We propose a robust framework that helps in better domain adaptation by reducing the probabilistic and subspace shift between the domains without compromising with their distributional information, and diminishing the distance of the between-class and within-class scatter of the domains collectively. The algorithm generates pseudo-labels after each iteration to update its objective function, thus helping it to perform better than conventional methods. The proposed framework tackles non-linear divergence by projecting the features into the kernel space. Computational experiments and their analysis show that the proposed algorithm performs better than other state-of-the-art domain adaptation methods on various visual recognition tasks.