Image Ranking with Density Trees for Google Maps

Abstract

We propose an unsupervised learning technique for image ranking of photos contributed by Google Maps users. A density tree is built for each point-of-interest (POI), such as The National Mall or the Louvre. This tree is used to construct clusters, which are then ranked based on size and quality. We choose a representative image for each cluster, resulting in a ranked set of high-quality, diverse, and relevant images for each POI. We validated our algorithm in a side-by-side preference study.