The team NERCMS ranked the first in 8 items in an international famous video retri uation competition

In the TRECVID2015 which is one of the most famous video retrieval evaluation competitions, the team of National Engineering Research Center for Multimedia Software (NERCMS) by the leadership of professor Ruimin Hu, the dean of computer science of Wuhan University, has done a good job in Instance Search Task. In 8 out of 30 objects retrieval provided by the organizer, this team was the NO.1 in the accuracy of retrieval.

Along with the popularity of the computer technical and the fast development of Internet application, users can get in touch with the increasing information of multimedia, such as the information of audio, video and images. How to retrieve the information which we really need fast and accurately in the large-scale of multimedia data has become an important problem affecting people’s lives, works and study. However, due to the intrinsic unstructured features of multimedia information (especially for audio, video and images), traditional way of text retrieval cannot meet the people’s needs of multimedia information retrieval. Different from traditional technical of text labels, the multimedia information retrieval technical based on content has provided a new approach to people. By searching and comparing the robust features multimedia content, people may find out the most needed information in the big multimedia data. Against this background, from the year of 2001, the National Institute of Standards and Technology (NIST) of America has held 15 competitions of TRECVID annually facing the video retrieval and evaluation. The TRECVID organizer provides the standard test data to the participating famous universities and scientific research institutions. The participants design their own algorithm according to these data and submit their results to the committee within the set time. These results will be evaluated and compared by the committee of NIST. The TRECVID2015 attracts 69 teams all over the world, including well-known universities and scientific research institutions in China and overseas, such as Carnegie Mellon University, IBM Research and so on.

The team NERCMS participated the Instance Search task in TRECVID2015. This task aims to retrieve 30 specific topics in the large-scale video database (TB magnitude). These topics include people, cars, and objects and so on. The different sizes and kinds of topics and the much noise in the background becomes a big test to the performance and effect of retrieval algorithm. Therefore the team NERCMS adopted a query adaptive method to measure the similarity between query and gallery images. With the help of cross-mode information such as text information, the team obtained an original retrieval results. After that, the strategies of Query Expansion with Adjacent Shots optimized the original results and improved the accuracy of algorithm. According to the official reports of NIST, the results submitted by the NERCMS has ranked the first place in the 8 out of 30 retrieval objects and The comprehensive result of our performance is Top 4 all over the world, which surpasses some famous research institutions, such as NTT from Japan, the UQ from Australia, the TUC from German and so on.

As the sole participants in Wuhan University, the team NERCMS has participated in this competition for the third time. Compared to the other competitors of TRECVID, we start late but progress fast. The retrieval accuracy of NERCMS this year is 0.367, which is 50 times higher than the first results in 2013. This progress has made NERCMS be in the first class with famous research institutions of multimedia information retrieval such as NII and CityU. The related technical has applied in objects retrieval in the surveillance system and played an important role in some cases, which is quite important for public safety.

The team members of NERCMS are all from the school of computer science of Wuhan University, including five graduate students (Lei Yao, Mang Ye, Jun Liu, Zheng Wang and Bingyue Huang) and two undergraduate students (Dongjing Liu and Tao Liu). The instructors are Pro. Ruimin Hu, Pro. Jun Chen and Chao Liang.