System Description: We have extracted 10 types of features including low level feature, mid-level semantic concepts feature. Multiple modalities features including visual, audio, text are applied in our framework. Non-linear SVM classifier is used to train the event model and generate the final event search engine results.
Metadata Generator Description: We have extracted 10 types of features including sparse SIFT, dense SIFT, color SIFT, GIST, PHOG, classemes, overfeat, OCR, ASR, MFCC.
Semantic Query Generator Description: For semantic query generation, given an event query, we measure the similarity between the query and the 3,658 concepts using ConceptNet and WordNet, and then rank the top ranked semantic concepts in our concept pool as the candidate semantic concepts.
Event Query Generator Description: We simply combine the semantic query with a basic event detector node that specifies the event id and exemplar set. 
Event Search Description: We have applied non-linear Chi-square kernel SVM to train the classifier and generate the event search engine.
Training data and knowledge sources:  Tesseract OCR, overfeat, classemes, conceptnet, and wordnet.