Stereotactic human brain atlases, either in print or electronic form, are useful not only in functional neurosurgery, but also in neuroradiology, human brain mapping, and neuroscience education. The existing atlases represent structures on 2D plates taken at variable, often large intervals, which limit their applications. To overcome this problem, we propose a hybrid interpolation approach to build high-resolution brain atlases from the existing ones. In this approach, all section regions of each object are grouped into two types of components: simple and complex. A NURBS-based method is designed for interpolation of the simple components, and a distance map-based method for the complex components. Once all individual objects in the atlas are interpolated, the results are combined hierarchically in a bottom-up manner to produce the interpolation of the entire atlas. In the procedure, different knowledge-based and heuristic strategies are used to preserve various topological relationships. The proposed approach has been validated quantitatively and used for interpolation of two stereotactic brain atlases: the Talairach-Tournoux atlas and Schaltenbrand-Wahren atlas. The interpolations produced are of high resolution and feature high accuracy, 3D consistency, smooth surface, and preserved topology. They potentially open new applications for electronic stereotactic brain atlases, such as atlas reformatting, accurate 3D display, and 3D nonlinear warping against normal and pathological scans. The proposed approach is also potentially useful in other applications, which require interpolation and 3D modeling from sparse and/or variable intersection interval data. An example of 3D modeling of an infarct from MR diffusion images is presented.