How many more images do we need? Performance prediction of bootstrapping for image classification