Segmenting Photo Streams in Events Based on Optical Metadata

Traditional methods for event segmentation of photo streams use time and/or content-based information. In this paper, we present event segmentation from a novel perspective. We propose to segment photo streams in events based on the scene brightness of photos by assuming that big scene brightness change implies an event transition of interest. The scene brightness is derived from camera parameters that are automatically set when photos are taken and recorded with each photo as metadata in standard forms like EXIF data. This information is available from metadata and is very inexpensive computationally resulting in fast segmentation. Hierarchical agglomerative clustering method is applied to build the event hierarchy of the photo stream based on the scene brightness difference. The proposed approach has been tested on several photo streams and very promising results have been obtained.