Search

People are increasingly becoming accustomed to taking photos of everyday life in modern cities and uploading them on major photo-sharing social media sites. These sites contain numerous pictures, but many have incomplete or blurred location information. The geo-localization of crowd-sourced pictures enriches the information contained therein, and is applicable to activities such as urban construction, urban landscape analysis, and crime tracking. However, geo-localization faces huge technical challenges. This paper proposes a method for large-scale geo-localization of crowd-sourced pictures. Our approach uses structured, organized Street View images as a reference dataset and employs a three-step strategy of coarse geo-localization by image retrieval, selecting reliable matches by image registration, and fine geo-localization by 3D reconstruction to attach geographic tags to pictures from unidentified sources. 3D reconstruction based on close-range photogrammetry is used to restore the 3D geographical information of the crowd-sourced pictures, resulting in the proposed method improving the median error from 256.7 m to 69.0 m, and the percentage of the geo-localized query pictures under a 50 m error requirement from 17.2% to 43.2% compared with the previous method. Another discovery of the proposed method is that, regarding the causes of reconstruction error, closer distances from the query cameras to the main objects in query pictures tend to produce smaller errors. The proposed method is not limited to small areas, and could be expanded to cities and larger areas owing to its flexible parameters.