Section 1 System Description We first extract visual feature for the videos. Every video is represented by a vector. Then we use SVM to train a model and score test videos. Section 2 Metadata Generator Description First, we select 100 frames from each video. Then we use the CNN model (caffe.berkeleyvision.org) to extract the feature of each frame. At last, we average the 100 image features as the metadata for the video. Section 3 Semantic Query Generator Description Our system does not generate semantic query. Section 4 Event Query Generator Description We use training data provided for each event to train a SVM classifier. Positive samples, near-miss samples and negative samples are given different C. Section 5 Event Search Description We use the probability output of SVM model as the result. Section 6 Training data and knowledge sources. We used the model provide by Caffe (http://caffe.berkeleyvision.org/getting_pretrained_models.html) for our metadata generation.