How to Train a CAT: Learning Canonical Appearance Transformations for Robust Direct Localization Under Illumination Change.