We discuss the accuracy limits for the localization of surfaces in three-dimensional (3-D) space. Such a localization is necessary for the registration of different views of an object, taken by 3-D sensors from several directions. A quantitative analysis shows that the lateral localization accuracy of a small surface area is proportional to the local curvature of the surface. This confirms the intuitive conjecture that our visual system performs localization of 3-D objects via sharp features. The longitudinal localization accuracy depends only on the noise of the data and is usually much better than the lateral localization accuracy, suggesting that surfaces are to be registered only along the longitudinal directions.