Scatter Matrix Normalization for Unsupervised Domain Adaptation

Abstract

The field of Domain Adaptation(DA) involves the usage of data from a source to train a model, and then predict the class of data samples of a different distribution. Domain Adaptation (DA) aims to leverage the available training and testing data to model a target domain classifier. Domain invariant features are extracted, and are used to minimize the distribution divergence between the source and target domains. The existing works do not consider reducing the discrepancy between the source and target covariance matrices, an important information source. No previous work has incorporated all objectives like manifold feature learning, scatter matrix normalization, discriminative information preservation, variance maximization, divergence minimization and geometric similarity preservation into a single objective function. In this work, we propose a novel domain adaptation framework for image classification that utilizes the covariance matrices of the source and target domains along with other important objectives like discrimination information preservation, divergence minimization, among others. A robust objective function that comprises of all these objectives is designed for optimal performance of the algorithm. The significance and impact of different types of normalization on the overall performance of the algorithm is also described. Experiments on benchmark domain adaptation datasets like PIE and Office-Home signify improvements over existing state of the art algorithms.