We focus on learning open-vocabulary visual classifiers, which scale up to a large portion of natural language vocabulary (e.g., over tens of thousands of classes). In particular, the training data ...