On the automatic online collection of training data for visual event modeling