Matching specific visual landmarks (like object corners) between a demonstration and a new scene to align actions.