Rock Painting , South Africa

s were quite different from those made directly by humans, the subjects could not tell which were better [8]. For browsing and searching large information archives, many users are familiar with Web interfaces. Therefore, our abstracting tool can compile its results into an html page, including the anchors for playing short video clips (see Figure 5). The top of the page in Figure 5 shows the film title, an animated gif image constructed from the text bitmaps (including the title), and the title sequence as a video clip. This information is followed by a randomly selected subset of special events, which are followed by a temporally ordered list of the scenes constructed by our shot-clustering algorithms. The bottom part of the page lists the creation parameters of the abstract, such as creation time, length, and statistics. Video abstracting is a young research field. We would like to mention two other systems suitable for creating abstracts of long videos. The first is video skimming [2], which mainly seeks to abstract documentaries and newscasts. It assumes that a transcript of the video is available; the video and the transcript are then aligned by word spotting. The audio track of the video skim is constructed by using language analysis to identify important words in the transcript; audio clips around those words are then cut out. Based on detected faces [10], text, and camera operation, video clips are selected from the surrounding frames. The second system is based on the image track only, generating not a video abstract but a static scene graph of thumbnail images on a 2D “canvas.” The scene graph represents the flow of the story in the form of keyframes, allowing users to interactively descend into the story by selecting a story unit of the graph [11]. Conclusions Using the algorithms discussed here for automatically generating video abstracts, we first decompose the input video into semantic units, called “shot clusters” or “scenes.” Then we detect and extract semantically rich pieces, especially text from the title sequence and special events, such as dialog, gunfire, and explosions. Video clips, audio clips, images, and text are extracted and composed into an abstract. The output can then be compiled into an html page for easy access through browsers. We expect our tools to be used for large multimedia archives in which video abstracts would be a much more powerful browsing technique than textual abstracts. For example, broadcast stations today sit on a gold mine of archived difficult-to-access video material. Another application of our technique could be to create an online TV guide on the Web, with short abstracts of upcoming shows, documentaries, and feature films. Just how well the generated abstracts capture the essentials of all kinds of videos remains to be seen in a larger series of practical experiments.