Automatic reconstruction of three-dimensional objects and environments from photographic images is important for many applications that integrate virtual and real data. Many approaches have been used to solve the problem. Traditional methods are based on matching features [34] or textures [14]. In traditional stereo methods, for a complete 3D model of real object, many partial models must be computed with respect to a set of base viewpoints, and the surface patches must be fused into a single consistent model. The procedure is called multi-view registration [4]. A popular method in this application is Iterative ClosestPoints(ICP) [31]. A parameterized model needs to be fitted to the sparse 3D points to obtain final dense surface reconstruction. Furthermore, in ICP there is no explicit handling of occlusion. Meshes and/or systems of particles [15,18] surface representations can be deformed according to constraints derived from images. These mesh and/or particle approaches may end up causing part of the mesh or particles to cluster in areas of high curvature, resulting in instabilities in subsequent deformation move; they usually fail to capture shapes with complicated topology. More recently, voxel-based approaches have been widely used to represent 3D shape [14,32,21,5,9].The computation is based on 3D scene space instead of image space. Marching-cube-like techniques [24] are still necessary to get the parameterized surface. The space carving method [21,22] aims at recovering a one-parameter family of volumes that are increasingly tighter supersets of the true scene. Once a voxel is carved away, it cannot be recovered, and any erroneous carving will propagate into further erroneous carvings. This is partially alleviated by probabilistic space carving [6]. The level set based method [14] is based upon variational analysis of the objects in the scene and their images. Its implicit representation can handle topology changes automatically. In order to overcome the large numerical complexity of implicit representation of objects in the level set approach, Ron Kimmel [20] presented a depth function method, which operates on a surface represented by a depth function (hence the surface is restricted to that function). The explicit representation of the surface makes it more efficient computationally, at the cost of being limited to surface patches. Working with a fixed coordinate system, a single surface function suffices instead of the 3D space in which the level set surface is embedded. There are also attempts to achieve multi-resolution using space carving and level sets methods [33]. The method starts with coarse settings that are refined as necessary. The method is a global approach, which means refinement occurs all over the currently recovered shape.