It has been debated whether object recognition depends on structural or view-specific representations. This issue is revisited here using a paradigm of priming, supervised category learning, and generalization to novel viewpoints. Results show that structural representations can be learned for three-dimensional (3D) objects lacking generalized-cone components (geons). Metric relations between object parts are distinctive features under such conditions. Representations preserving 3D structure are learned provided prior knowledge of object shape and sufficient image input information is available; otherwise view-specific representations are generated. These findings indicate that structural and view-specific representations are related through shifts of representation induced by learning.