Metric Feature Indexing for Interactive Multimedia Search

We address the problem of interactive event search in large video archives, using multimedia cues derived from both audio and visual information. Content-based search tools are limited in their ability to quickly respond to ad hoc queries, where a user defines an event of interest on the fly from several video clips. We show that an index organizing pre-computed features can be used to determine the highest-scoring detections while significantly reducing the number of clips examined, and that this approach is compatible with state of the art event search using late fusion over several multimedia cues. In experiments with the TRECVID Multimedia Event Detection (MEDTEST) data, we show that the top 100 detections can be determined by examining only about 1% of the video archive, and that time-limited mean average precision is significantly improved relative to a comparable approach from the literature.