I presume they use the image selection as training data too—if not that seems like awfully low hanging data fruit.