Multimedia Database Technologies for Sophisticated Analysis of Large-Scale Video Corpora

建立一个庞大的多媒体档案的成本已经大大降低。特别是，广播视频档案是有用的多媒体研究，因为它们涉及的内容广泛。人们自然地认为，大量的广播视频将是多媒体索引和挖掘研究的有用语料库。文本语料库被广泛用于文本处理研究。它们有助于文本处理技术的进步。对于多媒体信息处理，可能出现相同的现象。然而，一般来说，视频分析非常耗时。一台PC需要一个小时来分析一个小时的视频是很常见的。这意味着我们需要数百台PC在一天内处理数千小时的视频。在不久的将来，用数百个CPU核心处理数千个视频可能是常见的。因此，HPC和DB技术也将成为大规模多媒体信息处理的重要组成部分。从这个角度出发，我们对分析大规模视频语料库的多媒体数据库技术进行了初步的研究。我们从我们的视频档案中提取了大约550，000个镜头，这些镜头记录了七年（2，400天/1，200小时）的每日新闻节目。通过从每个相机镜头中选择一个代表性的帧，我们获得了包含各种图像的集合。然后，我们应用我们的索引技术，包括SR-Tree（一种多维索引结构）和DSNN（独特性敏感的最近邻搜索算法），来检查数据库技术的有效性。根据我们的性能实验，我们证实了数据库技术是非常有效的，以减少计算和I/O成本的多媒体索引。

The cost of building a huge multimedia archive has been significantly reduced. Especially, broadcast video archives are useful for multimedia research since they involve a wide variety of contents. It is naturally perceived that a huge amount of broadcast video would be a useful corpus for multimedia indexing and mining research. Text corpora are widely used for text processing research. They contribute to the advancement of text processing techniques. The same phenomenon may arise for multimedia information processing. However, in general, video analysis is exhaustively time consuming. It is quite common to take one hour for one PC to analyze one-hour video. This means that we need hundreds of PCs to process thousands hours of video in a day. In near future, it may be common to process thousands of videos with hundreds of CPU cores. Thus, HPC and DB Technologies would also be an important element for large-scale multimedia information processing. From this viewpoint, we conducted a preliminary study on multimedia database technologies for analyzing large-scale video corpora. We extracted about 550, 000 camera shots from our video archive which records a daily news program for seven years (2, 400 days/1, 200 hours). By choosing a representative frame from each camera shot, we obtained a collection containing wide variety of images. Then, we applied our indexing techniques, including SR-Tree (a multidimensional index structure) and DSNN (distinctiveness-sensitive nearest-neighbor search algorithm), to examine the effectiveness of database technologies. According to our performance experiments, we confirmed that database technologies are quite effective for reducing both computation and I/O costs of multimedia indexing.