Typical data acquired in physical experiments or in geometrical or medical imaging are discrete. This data is generally interpreted as samples of an unknown function, whose derivatives still serve for the data characterisation. For example, the movement of a fluid is described as a velocity field, a curve is characterised by the evolution of its curvature, images used in medical sciences are usually segmented by estimates of their gradients, among others. It is possible to obtain coherent derivatives by filtering the data. However, with multidimensional data, the usual filters present a bias towards to favor directions aligned with the axis, which may induce problems when the derivatives are interpreted geometrically. For example, the estimated curvature would depend on the orientation of the curve, loosing the geometric meaning of the curvature. The goal of the present work is to improve the geometric invariance of derivative filters.