Explicitly disentangling image content from translation and rotation with spatial-VAE