Domain adaptation by manifold transfer for image classification

There are many progress towards transfer learning tasks where the distribution of two datasets is different. Most of them try to build a classifier from source dataset and then apply it to target dataset to obtain more correct labels for target samples by transferring knowledge between two domains. But they pay no attention to manifold transfer across datasets. In this paper, our goal is to inject the structure of source data to target part and learn a mapping matrix that maps both datasets to a new subspace in which the target manifold follows the properties of source manifold. In the next stage, we learn a metric such that distinct domains get close to each other by decreasing the difference between their means through marginal and conditional adjustment. Then, our algorithm uses marginalized denoising and low-rank property to increase the robustness of trained metric. We evaluate the validity of our algorithm by testing it on 12 cross-domain image datasets and show its ability against other domain adaptation approaches in classification tasks.