The authors identify and describe problems with biases in data used to train current machine learning systems, introducing a crowdsourcing platform to college a large dataset of object from many different views. They also introduce ObjectNet, 40k crowdsourced images that can be used as a test set for object recognition with variation in object rotation, viewpoints, backgrounds. Evaluation demonstrates that the ImageNet dataset is not sufficient for learning models that are robust to these kinds of object variations. This is a strong paper with several strong contributions