Query by Shots: Content-based Video Retrieval Using Multiple Examples and Rough Set Theory

Since the same event (e.g. ``person opening a door'' and ``people with mostly vegetation in the background) is taken by various camera and editing techniques, shots showing the event have significantly different low-level features. Thus, the event cannot be internally defined by a single model. To overcome this, we propose an example-based event retrieval method which uses multiple examples to externally define the event. Specifically, given positive and negative examples for the event, we use ``rough set theory'' to extract subsets of the event. Here, in each subset, examples can be correctly defined by a different combination of low-level features. Like this, we define the event by multiple models to cover large variations of low-level features. The experimental results on TRECVID 2008 video data indicate a possibility of our method.