In this paper, we propose view-based recognition, a method for 3D~object recognition based on multi-view representations. We analyze view-based recognition and compare its performance theoretically and empirically with one of the most commonly used method for 3D~object recognition, 3D~bounded error recognition. In particular, we show that the probability of false positive or false negative matches in a view-based recognition system is not substantially different from the probability of similar errors in other commonly used recognition systems. Furthermore, we derive an upper bound on the number of views needed to be stored by a view-based recognition system in order to achieve zero probability of false negative matches. Simulations and experiments on real images suggest that these estimates are conservative and that view-based recognition is a robust and simple alternative to the more traditional 3D~shape based recognition methods.