Enhancing Person Annotation

Rapid technological growth and the decreasing cost of photo capture means that we are all taking more digital photographs than ever before. However, lack of technology for automatically organising personal photo archives has resulted in many users left with poorly annotated photos, causing them great frustration when such photo collections are to be browsed or searched at a later time. As a result, there has recently been significant research interest in technologies for supporting effective annotation. This thesis addresses an important sub-problem of the broad annotation problem, namely “person annotation” associated with personal digital photo management. Solutions to this problem are provided using content analysis tools in combination with context data within the experimental photo management framework, called “MediAssist”. Readily available image metadata, such as location and date/time, are captured from digital cameras with in-built GPS functionality, and thus provide knowledge about when and where the photos were taken. Such information is then used to identify the “real-world” events corresponding to certain activities in the photo capture process. The problem of enabling effective person annotation is formulated in such a way that both “within-event” and “cross-event” relationships of persons’ appearances are captured. The research reported in the thesis is built upon a firm foundation of content-based analysis technologies, namely face detection, face recognition, and body-patch matching together with data fusion. Two annotation models are investigated in this thesis, namely progressive and nonprogressive. The effectiveness of each model is evaluated against varying proportions of initial annotation, and the type of initial annotation based on individual and combined face, body-patch and person-context information sources. The results reported in the thesis strongly validate the use of multiple information sources for person annotation whilst emphasising the advantage of event-based photo analysis in real-life photo management systems.