Interactive Video Retrieval System Integrating Visual Search with Textual Search

Intelligent Multimedia Knowledge Management

Volume

Papers from the 2003 AAAI Spring Symposium

Contents

In this paper, we propose a technique for efficient retrieval of videos and scenes in large digital video archives. This technique has two features. The first involves displaying several videos at the same time so that users can see many videos in a short time. The second involves detecting cut frame images from a video automatically and arranging them in a virtual 3D space in which similar cut frame images are located close to each other by using self organization map (SOM). Users can search for scenes intuitively and efficiently by only having to look in an area where images that look like the target image are gathered. We conducted experiments to confirm the effectiveness of the scene retrieval technique. The results show that retrieval using this technique is at least twice as fast as retrieval using fast-forwarding.