Model transferring for unsupervised object localization

Object localization becomes considerable research with the rapid expansion of computer vision applications. How to locate an efficient object for vision recognition tasks is constantly a pivotal issue. Unsupervised object localization is extremely challenging because of the deletion of label information. In the last decade, based on a large amount of label information, a lot of localization algorithms have been put forward, some of which are capable of obtaining desirable performance. Whereas, it is unaffordable for providing sufficient labelled data in plenty of real-world applications, which is the reason why unsupervised object localization captured growing attention. In such cases, model transfer would dramatically improve the performance by evading much expensive labelling efforts. Towards this end, this paper focuses on model transferring for solving the unsupervised object localization problem. We propose an effective method for utilizing the pre-trained deep convolutional models to evaluate the correlation of features and predict results between original image and sub-images which generated via object proposal. The most similar sub-image can precisely locate the object for the unlabeled images. There is one more point, extensive experiments validate the effectiveness of the proposed method. On publicly available object localization datasets, this method invariably outperforms existing state-of-the-art methods.