Robust Two Stage Unsupervised Metric Learning for Domain Adaptation

Most commonly used metric learning procedures suppose that the input feature space and domain of the training and test data are identical. In such cases these algorithms cannot improve target learning problems. This paper presents a robust distance metric for domain adaptation in two stages. At first stage both source and target features are transferred to a newly found latent feature space, which minimizes the difference between domains as well as the data properties are preserved. Then in the second stage, the desired metric is learned with a marginalized denoising strategy and the low-rank constraint. To show the superiority and power of the proposed method it is tested on distinct kinds of cross-domain image categorization datasets and the results prove that our approach remarkably exceeds other existing domain adaptation algorithms in the classification tasks.