Capture-time Classification of Mobile Sunset Photos Leveraging Strong Spatiotemporal Cues
Ali Almajed, Matthew R. Boutell · 2014
Advancements in the field of mobile photography provide new contextual cues to enhance capture-time scene classification. In this paper we present a novel approach to improve an existing sunset photo classifier by using the photo's spatiotemporal cues. First, we classify the photo based on visual cues using a support vector machine. Second, we use spatiotemporal cues -- geolocation, date, and time -- to calculate the range of times when sunset photos are expected to be taken that day; the probability distribution is learnt from a large set of geotags obtained from Flickr. We then classify the photo using those spatiotemporal cues. Finally, we obtain a classification from the posterior probability using both visual and spatiotemporal cues. We present a new mobile camera app that classifies photos as sunset or non-sunset at capture time, and demonstrate the effectiveness of our application on a large dataset.