Unsupervised Learning of Visual Sense Models for Polysemous Words