We introduce a new method to reconstruct 3D objects using a set of volumetric primitives, i.e., superquadrics. The method hierarchically decomposes a target 3D object into pairs of superquadrics recovering finer and finer details. While such hierarchical methods have been studied before, we introduce a new way of splitting the object space using only properties of the predicted superquadrics. The method is trained and evaluated on the ShapeNet dataset. The results of our experiments suggest that reasonable reconstructions can be obtained with the proposed approach for a diverse set of objects with complex geometry.