In an International Conference, award-winning hard, in an area of top-level meeting winning more difficult, in a top-level international meetings for two awards are rarely precedent.
However, Microsoft Research Asia. At the end of September this year in Germany next to the city of Munich, Augsburg, ACM multimedia Conference (ACM Multimedia2007), we also won the best paper award and best demo award.
ACM Multimedia – multimedia areas of top-level event
ACM Multimedia is the most advanced in the field of multimedia session, participants are multimedia content analysis, multimedia applications, systems and interaction of researchers.
Meetings usually include long papers, short papers, presentations and video presentations, and other different plates, which long papers is the most challenging in recent years, every year, only about 50 to 60, the rate of 20 per cent of papers. Microsoft Research Asia since 2003 in this meeting has been very good performance, especially this year, the Institute has a total of eight articles long papers accepted, set a new high, representing the total number of papers (57 articles) of 14%. Which I take of the multimedia content analysis and search team has five is received (including University of science and technology cooperation), and I and his Ph.d interns Qi Jun and other four colleagues cooperation after the review of papers was elected to a four-story, one of the best thesis candidates will compete for the best paper award. (Stage winner — best paper award)
This year at the mainstream papers (approximately one third) is a multimedia search and its related applications, including video and image annotation, video search, video, advertising, media, recommended, and so on.
Research Institute of the eight articles long papers are associated with this topic. This shows that the Internet is booming on the great multimedia areas, as well as academia and industry are looking to media technology in Internet communication, share, search, recommendation and advertising applications. Video surveillance, mobile devices multimedia application, media, interactive, video transmission, etc. are hot topics.
This is also Institute to participate in the ACM Multimedia in the largest number of years, including twelve interns.
It was also met with several previous research intern, they now are in foreign universities to continue learning, and in this time the General Assembly also has an article published.
Best paper-I at first do report
Elected to the world's top academic conference best paper of candidates is not easy, and best paper of speech is the award-winning thesis is an important reference.
I had the privilege to act as representative of all authors, in the General Assembly of the first technical Session "best paper contest" in the first report, that is, the entire session of the first technical presentations. In Augsburg, in the members and the supervision of our team intern evaluation, we have a walkthrough times. Although slightly before, but confident. The Panel discussion, our work stand out. In the second day of up to five hours for dinner, we took the best paper award. (Conference of China first win, use the previous day in Augsburg city shooting pictures live demo)
Award-winning thesis is a video search and machine learning the latest work, the main research how to use semantic concepts (keywords) linkages to improve the accuracy of the automatic video annotations.
Video annotations are based on the content of the video search key steps, automatic dimensioning good video can use text search technology to index. There are many concept learning methods can be divided into two categories, the concept of independent processing, will be more conceptual learning into multiple classification in two classes. The disadvantage of this type of method is not using semantic concepts. In real-world problems, contact between the semantic concept, and that link can be used to improve the effectiveness of the callout, and you can even take advantage of easy-to-detect semantic concepts to help detect the concept is difficult to detect. The second method is based on the fusion of ways, but the integration is in the first class method based on, or through the integration of separate classification of output results to improve the accuracy of marking. Such methods use a semantic, but this kind of "two-step" policies does not resolve the problem of error propagation, sometimes even reduce annotation results. We propose a concept for both the semantics and the modeling of the relationship between the concept of the new method, called "multiple concepts associated learning" (CorrelativeMulti-Label Learning(CML)）。 This approach to overcome the shortcomings of existing methods, in large datasets on achieved gratifying results. Specific papers see: http://research.microsoft.com/~xshua/.
No comments:
Post a Comment