Acquiring human shape is a prerequisite to many applications in augmented and virtual reality, as well as in computer graphics and animation. The acquisition of a real person must be precise enough to have the best possible (realistic) rendering. To do so in real time, "Shape-from-silhouette" (SFS) methods are used. One limitation of these methods is that the acquired subject must be visible from all the camera filming it. If not, some parts of the object are not reconstructed in 3d. This paper presents a modified SFS algorithm, that extends the 3d reconstruction space. Our extension allows to build an estimation of an object's 3d shape even if it comes out of sight from one or more cameras.