Comparing curvature estimation techniques.

Abstract

This article presents a careful comparative evaluation of two techniques for numerical curvature estimation of 2D closed contours (more specifically closed, regular and simple parametric curves). The considered methods are: (a) a 1-D Fourier-based approach; and (b) a 2-D Fourier-based approach involving the embedding of the contour into a 2-D regular surface (presented for the first time in this article). Both these techniques employ Gaussian smoothing as a regularizing condition in order to estimate the first and second derivatives needed for curvature estimation. These methods are considered according to a multiresolution approach, where the standard deviation of the Gaussians are used as scale parameters. The methods are applied to a standard set of curves whose analytical curvatures are known in order to estimate and compare the errors of the numerical approaches. Three kinds of parametric curves are considered: (i) curves with analytical description; (ii) curves synthesized in terms of Fourier components of curvature; and (iii) curves obtained by splines. A precise comparison methodology is devised which includes the adoption of a common spatial quantization approach (namely square box quantization) and the explicit consideration of the influence of the related smoothing parameters. The obtained results indicate that the 1- D approach is not only faster, but also more accurate. However, the 2-D approach is still interesting and reasonably accurate for applications in situations where the curvature along the whole 2-D domains is needed.