1) System Description Our MED system combines following features: + Feature Extraction: Dense Trajectories (MBH descriptor), SIFT (Hessian Affine), MFCC + Feature Representation: Fisher Vector, GMM 256 + Training/Testing: Linear SVM 2) Metadata Generator Description + Dense Trajectories http://lear.inrialpes.fr/~wang/dense_trajectories. Only MBH is used for trajectory description + We extract SIFT using VLFeat with Hessian feature detector (http://www.vlfeat.org/overview/covdet.html) + MFCC: we use the library provided here http://labrosa.ee.columbia.edu/matlab/rastamat/ To saving the storage, we extract and encode raw features using Fisher Vector encoding on the fly, rather than storing raw features. The library for Fisher encoding is provided here http://www.robots.ox.ac.uk/~vgg/research/encoding_eval/. Hardware Usage: + Grid computer with 448 nodes, 4GB RAM each + 2 Servers with 24 cores and 128MB RAM each + 1 Server with 40 cores and 512MB RAM 3) Semantic Query Generator Description We don't use semantic information at the moment. 4) Event Query Generator Description We use LibSVM (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) with linear kernel for training. Cross-validation are used to tune learning parameters. The pre-computed kernel technique is employed to saving the computational time for different Event kit (i.e, EK10Ex, EK100Ex). Hardware Usage: + 1 Server with 40 cores and 512MB RAM 5) Event Search Description We use LibSVM (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) to predict the detection score of each video for each event category. Hardware Usage: + 1 Server with 40 cores and 512MB RAM 6) Training data and knowledge sources We only use data provided by TRECVID at the moment.