IM2MAP

Abstract

This paper investigates georeferenced social multimedia for geographic discovery. We propose a novel framework wherein large collections of community contributed photo collections are used to map phenomena not easily observable through other means. We employ a regression framework in which a limited number of labeled training images are used to learn a regressor. This regressor is then applied to large collections of novel images whose locations are known and the predictions are used to create maps.We propose two novel extensions to a standard regression approach. In the first, a graph Laplacian semi-supervised learning approach leverages unlabeled images to improve the accuracy of the regressor. This is important because it allows us to exploit large collections of community contributed photos while limiting the number of images that need to be manually labeled. In the second extension, the regressor is based on a novel composite visual-geographic location kernel which considers both the visual characteristics and the geographic locations of images.We apply our approach to predict the scenicness of geographic locations at the country-scale based on the ground-level photos at the locations. While our results are noisy, this preliminary investigation demonstrates the feasibility of geographic discovery from georeferenced social media as well as the advantages provided by our extensions to a standard regression approach.