SIFT localisation accuracy on interpolated speckle images

Although sine interpolation is considered ideal for adequately sampled bandlimited data, speckle images are not always well sampled. This paper explores the relationship between sinc interpolation and localisation accuracy of SIFT features, as well as the effect of oversampling on the quality of feature correspondences. Since SIFT detects features in the Gaussian scale space, it is not concordant with sinc interpolation. The repeatability and localisation errors between images shifted by fractional displacements are demonstrated under the conditions of: no oversampling, two-times oversampling, and four-times oversampling. Feature matching is shown to be largely infeasible with only Nyquist sampling, whereas four-times oversampling significantly increases the number of detected features and the proportion of inlier matches, while the mean localisation error of these matches also decreases more than four-fold.