It is particularly important for high levels of the pyramid, where sample spacings correspond to large distances in the base image. In addition to increasing the accuracy of transformation computations, accurate interest point localisation also enables more accurate feature descriptors, which improves matching.