When objects cannot be represented well by single feature vectors, a collection of feature vectors can be used. This is what is done in Multiple Instance learning, where it is called a bag of instance
