The publication shows a strategy for the automatic geometric registration of two voxel datasets. The algorithm starts with the extraction of 3D point features in both datasets, which are obtained by binarization using deep-learning and connected components analysis. This is followed by the computation of a descriptor vector for each feature using its nearest neighbors. These descriptors are then used to determine corresponding features in both voxel datasets. Last, a 3D similarity transformation is computed to obtain the seven transformation parameters between the datasets. Herein, outliers in the determination of feature correspondences can be removed using Random Sample Consensus (RANSAC). The workflow will be demonstrated on an example of a voxel dataset pair from micro tomography scans of a concrete sample with air pores.