| Over the last few years, the advances in technology of digital cameras equipped with GPS and cameras on smart phones have resulted in high rate of generating digital multimedia data. Events in our lives are witnessed by such digital content. A semantic gap exists between media and people's perceptions of the events and memories associated with them. Since photos are the most ubiquitous and prolific form of user generated content, we focus on the automatic annotation of personal photo collection. This dissertation describes a framework that labels personal photo collections with events and subevents. In this thesis, we developed an automatic approach that assigns expressive tags to individual photos in a photo stream of an event. Given a domain event-model, our algorithm uses a new clustering method, abduction strategy, and a process that is called Ontology-Augmentation to obtain, refine, and validate expressive tags. We introduce a measure of plausibility for ranking and selecting the best possible subevent explanation for an image or a group of images. Our technique uses a combination of event ontology, contextual data sources on the web, image metadata (timestamp, location, and camera parameters), and image content for this task. We tested our approach on 27 personal photo albums with total number of 2253 images. |
