Predicting Event Memorability from Contextual Visual Semantics