A method for estimating the oriented rose of normal directions of a three dimensional set $Z$ from a digitization of $Z$, i.e. a voxel image, is presented. It is based on counts of informative configurations in $n \times n \times n$ voxel cubes. An algorithm for finding all informative configurations is proposed and an estimation procedure is described in detail for the case $n=2$. The presented method is a 3D version of a method of estimating the oriented rose of binary planar images using $n \times n$ configurations. A new feature is the design-based approach, being more appropriate for biomedical image analysis than the formerly applied model-based approach.