Rapid deformations in snow are mainly controlled by particle rearrangements and contact interactions. To study this deformation regime, the description of the snow microstructure in terms of grains, which could eventually be handled by discrete element models, is relevant. In practice, microtomography has become a standard method to image the three-dimensional distribution of ice and pores, as a set of binary voxels. Here, we propose a new method to directly identify individual snow grains defined as zones separated by regions of potential mechanical weakness, in the microtomographic images. In general, these grains are not well separated but rather sintered together. Our new method, based on local geometrical criteria, is shown to detect contacts directly inferred from an explicit numerical mechanical experiment. The developed algorithm is tested on snow but is generic and applicable to various geomaterials with a granular-like microstructure.