Clustering geo-tagged photo collections using dynamic programming.

This paper describes methods for clustering photos that possess both time stamps and geographical coordinates as metadata. We present a two part method that first analyzes photos' time and location information to independently partition the photos into multiple clusterings. A subset of the detected clusters is then selected for the final photo clustering using an efficient dynamic programming procedure that optimizes a clustering fitness score. We propose fitness measures to produce clusterings that are coherent in space, time, or both. One group of scores directly measures within-cluster inter-photo distances. A second set of scores measures clusters' consistency with the reference clusterings. We present experiments that validate our method using multiple data sets.