Publication

Abstract : Domain adaptation (DA) is an efficient approach that transfers the knowledge gained from the source domain to learn an accurate classifier for the target domain. But the majority of the current domain adaptation algorithms assume that both the training and testing data have similar feature space with unaligned data distribution. Such algorithms fail when the feature representations vary. Heterogeneous Domain Adaptation (HDA) effectively addresses multi-modal data analysis where the features and distribution of source and target domain data are different. But, some of the existing heterogeneous domain adaptation approaches utilize either statistical or geometrical data alignment. Hence, we propose Joint Cross-Domain Preserving and Distribution Adaptation (JCDDA) for semi-supervised HDA, which learns a new latent subspace where the data samples from the source and target domain are projected. JCDDA preserves the cross-domain structure of the data and minimizes the discrimination differences between the domains. As our approach is semi-supervised HDA, we exploit the selective pseudo-labeling method to utilize unlabelled target data. We did extensive experiments using two benchmark object recognition datasets, and from our results, it is evident that our algorithm outperforms all of them.