Section 1 System Description
For the MED14 evaluation, we adopt various low level visual and audio features, as well as high level semantic concepts in our system. Here, we additionally collect 239 semantic concepts as a compliment to the 1,000 semantic concepts trained on ImageNet. In addition, we train different classifiers for these features and estimate fusion parameters on the development set.
 
Section 2 Metadata Generator Description
Given the video list downloaded from the I/O server, we extracted DenseSIFT, ColorSIFT, Improved Dense Trajectories, MFCC and concept scores for each video in the list.

Section 3 Semantic Query Generator
No semantic query is generated.

Section 4 Event Query Generator
We train SVMs classifiers with videos specified in the Exemplar Labels list given by the I/O server. Then different classifiers are combined with parameters learned on the development set.

Section 5 Event Search 
For each video in the EvalFull set, we apply the model trained in the Event Query Generator stage to obtain its final prediction score under the noPRF setting.

Section 6 Training data and knowledge sources
In addition to MED training corpora, on which the event models were trained, we adopted ImageNet and additional videos from YouTube to train concept detectors. Based on the concept detector scores as features, the event models were only trained on the provided MED corpora.