Rare is Interesting: Connecting Spatio-Temporal Behavior Patterns with Subjective Image Appeal

We analyze behavior patterns and photographic habits of the Nokia Mobile Data Challenge (NMDC) participants using GPS and timestamp data. We show that these patterns and habits can be used to estimate image appeal ratings of geotagged Flickr images. In order to do this, we summarize the behavior patterns of the individual NMDC participants into rare and repeating events using GPS coordinates and time stamps. We then retrieve, based on both the time and location information from these events, geotagged images and their "view" and "favorite" counts from Flickr. The appeal of an image is calculated as the ratio of favorite count to view count. We analyze how rare and repeating events are related to the appeal of the downloaded Flickr images and find that image appeal ratings are higher for events when the NMDC participants also took pictures and also higher for rare events. We thus design new event-based features to rate and rank the geotagged Flickr images. We measure the ranking performance of our algorithm by using the Flickr appeal ratings as ground truth. We show that our event-based features outperform visual-only features, which were previously used in image appeal ratings, and obtain a Spearman’s correlation coefficient of 0.47.