In this paper, we present a novel method to reconstruct the large scale scenes from multiple calibrated images. It first generates a quasi-dense 3D point cloud of the scene by matching key points across images. Then it builds a tetrahedral decomposition of space by computing the 3D Delaunay triangulation of the 3D point set. Finally, a triangular mesh of the scene is extracted by labeling Delaunay tetrahedra as inside or outside. A globally optimal label assignment is efficiently found as a minimum cut solution in a graph. Experimental results demonstrate the effectiveness of the proposed algorithm.