MKLab participated again this year in the annual TRECVID benchmarking activity, organized by the US National Institute for Standards and Technology (NIST). MKLab participated in four TRECVID tasks: Semantic Indexing (SIN), Event Detection in Internet Multimedia (MED), Multimedia Event Recounting (MER), and Instance Search (INS). Our results in these tasks were very good, significantly improving the results we had obtained in previous years.
Particularly for our participation in the Event Detection in Internet Multimedia (MED) task, we used a new, very fast machine learning method for big data problems, which we developed. This is a generic learning method, applicable to a wide range of problems that involve learning from big data. In large-scale multimedia event detection problems, it is shown to produce better results than both Kernel- and Linear-SVMs, while its training is one or two orders of magnitude faster than that of Linear-SVMs. In the TRECVID 2014 workshop, which took place in Orlando, Florida (USA), we gave a talk (http://www-nlpir.nist.gov/projects/tvpubs/tv14.slides/iti-certh.tv14.med.slides.pdf) on the use of this new machine learning method for event detection.