This report describes a technique to detect curvature. The technique uses local polynomial fitting on a local orientation description of an image. The idea is based on the theory of rotational symmetries which describes curvature, circles, star-patterns etc. The local polynomial fitting is shown to be equivalent to calculating partial derivatives on a lowpass version of the local orientation. The new method can therefore be very efficiently implemented both in the singlescale case and in the multiscale case.