Where was that photo taken? : deriving geographical information from image collections based on temporal exposure attributes

Abstract

This paper demonstrates a novel strategy for inferring approximate geographical information from the exposure information and temporal patterns of outdoor images in image collections. Image exposure is reliant on light and most photographs are therefore taken in daylight which again depends on the position of the sun. Clearly, the sun results in different lighting conditions at different geographical location and at different times of the day, and hence the observed intensity patterns can be used to deduce the approximate location of the photographer at the time the photographs were taken. Images taken inside or at night are temporally connected to the daylight images and the geographical information can therefore be transferred to related ‘‘dark’’ photographs. The strategy is efficient as it only considers meta information and not image contents. Large databases can therefore be indexed efficiently. Experimental results demonstrate that the current approach yields a longitudinal error of 15.7 and a latitudinal error of 30.5 for authentic image collections comprising a mixture of outdoor and indoor images. The strategy determined the correct hemisphere in all the tests. Although not as accurate as GPS receiver, the geographical information is sufficiently detailed to be useful. Applications include improved image retrieval, image browsing and automatic image tagging. The strategy does not require a GPS receiver and can be applied to the existing digital image collections.