Detecting Significant Events in Personal Image Collections

The organization and retrieval of images and videos is a problem for the typical consumer. A typical image collection includes many pictures of common activities that are not considered to be important by the user. These images inflate the number of assets in a collection to the point where it is difficult to find significant events when browsing. It is useful for the user to be able to browse an overview of important events in their collection. This paper proposes a new approach for identifying a small sub-set of events in a large collection that have a high probability of being significant. Using techniques from time-series modeling, a representation of a user's picture-taking behavior is constructed. The detection of significant events is based on the deviation from this learned representation. The results match a user's judgment of significance and enables efficient browsing and searching of the collection by focusing on a small set of images.