I. Introduction
Cross-domain object recognition is an important emerging topic in computer vision as the images used for training and testing often have different feature distribution. However, it is difficult to predict what will change between two domains. As far as we know, several unpredictable factors can affect the feature extraction and encoding results of an image, such as sensor character, object pose and location, image resolution, viewpoint, and background clutter. Therefore, in practice it is difficult to design domain-invariant descriptors. Recent study has shown that the feature distribution shift problem can also be caused by the data set bias problem [1], [2]. Because of the shift of distribution between two different domains, the performance of the classifier trained on the source domain tends to significantly degrade when testing on the target domain.