Cross-Domain Object Recognition Using Object Alignment

One popular solution to the problem of cross-domain object recognition is minimizing the difference between source and target distributions. Existing methods are devoted to minimizing that domain difference in a complex image space, which makes the problem hard to solve because of background influence. To discount the influence, we propose to minimize that difference using object alignment. We firstly present an algorithm to effectively align the object that appears in a set of images, and learn detectors for the aligned objects so that the detectors are robust to the influence of irrelevant background. Then we utilize the classification information from the image space to enhance our detectors. Finally, based on the detectors, we introduce a self-paced adaptation method to further reduce the domain difference. Experimental results demonstrate that the object alignment is effective to minimize the domain difference, and show the state-of-the-art recognition performance on several visual domain adaptation datasets.