An Automated Ground-to-Aerial Viewpoint Localization for Content Verification

In many forensic examinations and media authenticity verifications, it is essential to reconstruct the place where a photo was taken without having information such as GPS coordinates and other metadata available. In recent years, satellite imagery has been used in some of these reconstructions to map the image captured from the ground with the photos from above. This task, known as ground-to-aerial mapping, allows a very effective localization but remains a complex and time-consuming task. Deep learning-based methods allow for accurate automatic image matching; however, many of these solutions can be challenging to explain and, therefore, barely applicable in scenarios where it is necessary to justify the analysis. Consequently, in this paper, we propose a fully automated, explicable solution that is able to perform image-matching tasks based on a graph-based solution. Our proposed pipeline is composed of four stages in which we extract a graph representation of the images that we use for matching. Moreover, the designed pipeline improves previous related methods of 17.84% for the mean IOU top-1 and of 32.71% for the top-3.