This paper presents a new combinatorial approach to surface reconstruction with sharp features. Different from other postprocessing methods, the proposed method provides a systematic way to identify and reconstruct sharp features from unorganized sample points in one integrated reconstruction process. In addition, unlike other approximation methods, the reconstructed triangulated surface is guaranteed to pass through the original sample points. In this paper, the sample points in the sharp regions are defined as characteristic vertices (c-vertices), and their associated poles (c-poles) are used as a "sculptor" to extract triangles from a Delaunay structure for the sharp features. But, for smooth surface regions, an efficient region-growing scheme is used for triangle extraction and connection. Since only the c-poles associated with the sharp regions are used to participate in the Delaunay computation with the sample points, the proposed algorithm is adaptive in the sense that, given a sampled object with less sharp features, the triangulation becomes more efficient. To validate the proposed algorithm, some detailed illustrations are given. Experimental results show that it is robust and highly efficient.