This paper presents a new surface reconstruction method that extends previous carving methods for non-Lambertian objects by integrating the smoothness and image information of different aspects. We introduce a robust multi-view photo-consistency function and a single-view visual-consistency function. The former considers a general specularity that may introduce both reflectance models and robust statistics, while the latter is based on pixel homogeneity and continuity in individual images. The new consistent shape, with both multi-view and single-view, extends previous photo hull and visual hull, and is proven to be included in both photo hull and visual hull. This approximate shape is then refined by a graph-cut optimization method. Sample reconstructions from real sequences for both global and local optimizations are demonstrated.