The docs suggest creating a fake tag for images with nothing in them. This is going to take space in the model and try for it to optimize something which doesn't make sense. When I get an image with a false positive I want to be able to upload the image to the training set even though there is nothing of interest in the picture.