To train the models, they created a new dataset of synthetic images. These images show objects such as tea kettles or ...