TRECVid 2013 Semantic Video Concept Detection by NTT-MD-DUT

In this report, we describe the approaches and experiments on TRECVid 2013 video concept detection conducted by NTT Media Intelligence Laboratories in collaboration with Dalian University of Technology. For this year’s task, we focused our efforts on two aspects. For the first aspect, we investigated the state-of-the-art machine learning algorithm and feature representation for large-scale concept classifiers construction. Specifically, we first evaluated a newly developed powerful image representation which has been successfully adopted in other visual classification task, i.e., Fisher Vector, for concept detection. Meanwhile, we are also interested in the using of deep learning technique for video classification, and to this end, we have tested various settings of deep learning and the results are reported. For the second aspect, we followed the subspace partition based framework we proposed in our last year work and to balance the precision and efficiency, we proposed a sparse soft-clustering method for ensemble learning, which can get the optimal replication parameter. We conducted experiments on TRECVid SIN task evaluation dataset and submitted 4 runs based on the above methods.