Where The Photos Were Taken: Location Prediction By Learning From Flickr Photose

In this chapter, we explore the characteristics of geographically tagged Internet photos and determine their location based on the visual content. We develop a principled machine learning model to estimate geographical locations of photos by modeling the relationship between location and the photo content. To build reliable geographical estimators, it is important to find distinguishable geographical clusters in the world. These clusters cover general geographical regions not limited to just landmarks. Geographical clusters provide more training samples and hence lead to better recognition accuracy. We develop a framework for geographical cluster estimation, and employ latent variables to estimate the geographical clusters. To solve this estimation problem, we propose to build an efficient solver to find the latent clusters. We illustrate detailed qualitative results obtained from beaches photos taken at different continents. In addition, we show significantly improved quantitative results over other approaches for recognizing different beaches using the Flickr beach dataset as validation.