Balanced Weight Joint Geometrical and Statistical Alignment for Unsupervised Domain Adaptation

In real-world applications, images taken from different cameras usually have different resolution, illumination, poses, and background views. This problem leads to the need of domain adaptation in which case, training and testing are not drawn from the same distribution. There have been many studies carried out on domain adaptation, and among the state-of-the-art methods is the Joint Geometrical and Statistical Alignment (JGSA) approach. This paper presents an improvement for unsupervised domain adaptation in transfer learning using a Balanced Weight JGSA (BW-JGSA). The existing method of JGSA seeking the way to minimize the distribution divergence between marginal and conditional distribution across domains; however, treat them equally in terms of distribution weight. This drawback affects the existing method mainly when applied to real applications. The contribution of this paper is to use balanced distribution adaptation in JGSA that can adaptively leverage the importance of marginal and conditional distribution in JGSA. In this method, the balance weight factor, mu, will be applied to marginal and conditional distributions distance for each different subspace in JGSA. Comparing the proposed method with state-of-the-art techniques in object and digital datasets shows significant improvement of our work.