Gabriele Trivigno

Photo geolocation is a challenging task since many photos offer only few cues about their location. For instance, an image of a beach could be taken on many coasts across the world. Even when landmarks are present there can still be ambiguity: a photo of the Rialto Bridge could be taken either at its original location in Venice, or in Las Vegas. Traditional computer vision algorithms rely on the features provided to them during training. While this can lead to some degree of success when the data distribution of training and test data are the same, the problem becomes arduous when there is a domain shift between the two distributions. The goal of this PhD is to study the problem of visual geo-localization across visual domains. We will leverage over the intrinsic spatial connotation of place images and combine attention mechanisms with modern domain adaptation algorithms, in order to obtain perceptual representations that can be used for visual place recognition, as well as for content based image retrieval, able to close the domain gap differently on different parts of the images. Experiments will be conducted on publicly available databases as well as on data collections created during the project.