国際シンポジウム ～ビッグデータ時代における多メディア社会分析技術とその展望～

Abstract:
This talk will describe our recent work to develop methods and tools to handle multimedia queries, in order to identify videos as answers with multimedia justifications. Our specific goal is to extract, validate and present evidence extracted by automatic means from a large archive of video material collected from the Internet, to answer a video information need. The information need is specified in the form of an overall text description of the multimedia event to be found, a list of detailed components that make up the event such as objects, scenes, and activities, and finally some video examples of the type of video sought.
The core approach involves careful analysis or numerous low-level features, clustering, semantic concept analysis and finally fusion of all component results into a final retrieval ranking. The results of our CMU E-LAMP system submission to the NIST TRECVID evaluations of multimedia event detection and multimedia event recounting showed that our approach is quite effective at this task. The broader impact of this research lies in the ability to robustly analyze video of varying quality, regardless of origin in order to extract meaningful content. Ultimately, our grand vision is to extend our knowledge of computer vision in terms of how to 'understand' video content and enable video to be as searchable and accessible as text information is today.

Abstract:
The web represents a unique source of information that a growing number of scientific communities, agencies and industries are starting to need to mine at large scale. The ability to acquire, process and mine large scale of data from the web is becoming a strategic advantage in many domains from business intelligence to epidemiological tracking and monitoring.

This ability is already proving to be challenging for many actors as the infrastructure, the cost and the skills required constitute a significant barrier to entry.

But when it comes to doing this through time, all but a few (mainly large search engines) can do it at all. Yet, this is where the need is growing as it is about tracking trends and evolution, emergence and wide-spreading of anything that produce a signal or a trace on the web.

From this perspective, large scale archive of the web is to be considered as a strategic infrastructure of tomorrow's research, that will enable not only the study, from a sociological or political point of view of the evolution of society, but also of many more aspect of evolution, from low noise signal of ecological evolution, to economical trends, from emergence of new research of new research fields to tracking of reputation on the web etc.

A brief history of international efforts in this domain, challenges and approaches will be presented followed by a discussion on perspective for the future.