Viewpoint generalization in object recognition is the process that allows recognition of a given 3D object from many different viewpoints despite variations in its 2D projections. We used the canonical view effects as a foundation to empirically test the validity of a major theory in object recognition, the view-approximation model (Poggio & Edelman, 1990). This model predicts that generalization should be better when an object is first seen from a non-canonical view and then a canonical view than when seen in the reversed order. We also manipulated object similarity to study the degree to which this view generalization was constrained by shape details and task instructions (object vs. image recognition). Old-new recognition performance for basic and subordinate level objects was measured in separate blocks. We found that for object recognition, view generalization between canonical and non-canonical views was comparable for basic level objects. For subordinate level objects, recognition performance was more accurate from non-canonical to canonical views than the other way around. When the task was changed from object recognition to image recognition, the pattern of the results reversed. Interestingly, participants responded "old" to "new" images of "old" objects with a substantially higher rate than to "new" objects, despite instructions to the contrary, thereby indicating involuntary view generalization. Our empirical findings are incompatible with the prediction of the view-approximation theory, and argue against the hypothesis that views are stored independently.