Normal maps vs. visible images: Comparing classifiers and combining modalities