Ranklets are a family of multiscale, orientation-selective rank features suitable for characterising complex patterns. On square pixel lattices, ranklets bear a strong similarity to Haar wavelets. This extends to the sensitivity to horizontal and vertical edges. We propose a generalisation of ranklets to hexagonal pixels. The sixfold rotational symmetry of the lattice translates into features that are tuned to three preferential directions in the image. We present experimental results on a pattern recognition task (face detection) over a large image database, using both square and hexagonal pixels. Comparison of the results in the two cases confirms the consistent performanceof the generalised ranklets.