Localizing web videos using social images

While inferring the geo-locations of web images has been widely studied, there is limited work engaging in geo-location inference of web videos due to inadequate labeled samples available for training. However, such a geographical localization functionality is of great importance to help existing video sharing websites provide location-aware services, such as location-based video browsing, video geo-tag recommendation, and location sensitive video search on mobile devices. In this paper, we address the problem of localizing web videos through transferring large-scale web images with geographic tags to web videos, where near-duplicate detection between images and video frames is conducted to link the visually relevant web images and videos. To perform our approach, we choose the trustworthy web images by evaluating the consistency between the visual features and associated metadata of the collected images, therefore eliminating the noisy images. In doing so, a novel transfer learning algorithm is proposed to align the landmark prototypes across both domains of images and video frames, leading to a reliable prediction of the geo-locations of web videos. A group of experiments are carried out on two datasets which collect Flickr images and YouTube videos crawled from the Web. The experimental results demonstrate the effectiveness of our video geo-location inference approach which outperforms several competing approaches using the traditional frame-level video geo-location inference.