Hidden Markov model for event photo stream segmentation

A photo stream is a chronological sequence of photos. Most existing photo stream segmentation methods assume that a photo stream comprises of photos from multiple events and their goal is to produce groups of photos, each corresponding to an event, i.e. they perform automatic albuming. Even if these photos are grouped by event, sifting through the abundance of photos in each event is cumbersome. To help make photos of each event more manageable, we propose a photo stream segmentation method for an event photo stream - the chronological sequence of photos of a single event - to produce groups of photos, each corresponding to a photo-worthy moment in the event. Our method is based on a hidden Markov model with parameters learned from time, EXIF metadata, and visual information from 1) training data of unlabelled, unsegmented event photo streams and 2) the event photo stream we want to segment. In an experiment with over 5000 photos from 28 personal photo sets, our method outperformed all six baselines with statistical significance (p < 0.10 with the best baseline and p < 0.005 with the others).