Keele Research Repository

In this paper we describe our approach for the ImageCLEF-lifelog summarization task. A total of ten runs were submitted, which used only visual features, only metadata information, or both. In the first step, a set of relevant frames are drawn from the whole lifelog. Such frames must be of good visual quality, and match the given task semantically. For the automatic runs, this subset of images is clustered into events, and the key-frames are selected from the clusters iteratively. In the interactive runs, the user can select which frames to keep or discard in each interaction, and the clustering is adapted accordingly. We observe that the more relevant features to be used depend on the context and the nature of the input lifelog.