VideoClarity: High-speed meaning extraction in large video datasets

视频内容是蓬勃发展的英国和全球创意产业的主要商业资产，在当今的“大数据”洪流中占了很大比例。除了创意产业，视频也是一种主要的通信工具，占据了当今互联网流量的60%。然而，由于每周有数百万用户创建、处理和（不总是合法地）上传EB级的视频内容，视频仍然是大数据生态系统中最难管理的元素。这是因为目前所有用于视频中的高级语义描述的方法都需要手动注释或计算密集型视频解码和处理。因此，提供具有成本效益和有意义的视频搜索已被证明是一个无法克服的问题。这个项目有一个新的见解：编码相关元数据的隐藏源已经存在于现代压缩视频文件容器中，它足以用于自动标记和可视化。利用这种“提示”元数据可以使视频流的分析速度提高3~6个数量级，并降低成本，使EB级视频数据集能够在商品硬件上被新发现和分析。

Video content is a primary business asset of the thriving UK and global creative industries, and represents a substantial proportion of today’s “big data” deluge. Beyond the creative industries, video comprises a major communications tool, occupying 60% of today’s Internet traffic. However, with millions of users creating, processing and (not always legally) uploading exabytes of video content each week, video remains the least-manageable element of the big data ecosystem. This is because all current methods for high-level semantic description in video require either manual annotating or compute-intensive video decoding & processing. Delivering cost-effective and meaningful video search has therefore proved to be an insurmountable problem. This project takes a novel insight: that a hidden source of coding-related metadata already exists within modern compressed video file containers, and it is sufficient for automated tagging and visualization. Tapping into this “hint” metadata enables video streams to be analyzed with 3~6 orders increase in speed and decrease in cost, enabling exabyte-scale video datasets to be newly-discovered and analysed over commodity hardware.