In this paper, we describe the TRECVid 2012 videoconcept detection system first developed at the NTTMedia Intelligence Laboratories in collaborationwith Dalian University of Technology. For thisyear’s task, we adopted a subspace partition basedscheme for classifier learning, with emphasis on thereduction of classifier complexity, aiming atimproving the training efficiency and boosting theclassifier performance. As the video corpus used forTRECVid evaluation is ever increasing, two practicalissues are becoming more and more challenging forbuilding concept detection systems. The first one isthe time-consuming training and testing procedures,which have taken up most of the evaluation activities,preventing the design and testing of novel algorithms.The second and the more important issue is thatwhen using whole data for classifier training, thederived separating hyperplanes would be rathercomplex and thus degrade the classificationperformance. To address these issues, we propose toadopt the “divide-and-conquer” strategy for conceptdetector construction as follows. We first partitionthe whole training feature space into multiplesub-space with a scalable clustering method, andthen build sub-classifiers on these sub-spacesseparately for each concept. The decision of a testingsample is the fusion of the results a few firedsub-classifiers. Experimental results demonstrate theefficiency and effectiveness of our proposedapproach.