Shape from silhouette methods are extensively used to model dynamic and non-rigid objects using binary foreground-background images. Since the problem of reconstructing shapes from silhouettes is ambiguous, a number of solutions exist and several approaches only consider the one with a maximal volume, called the visual hull. However, the visual hull is not always a good approximation of shapes, in particular when observing smooth surfaces with few cameras. In this paper, we consider instead a class of solutions to the silhouette reconstruction problem that we call visual shapes. Such a class includes the visual hull, but also better approximations of the observed shapes which can take into account local assumptions such as smoothness, among others. Our contributions with respect to existing works is first to identify silhouette consistent shapes different from the visual hull, and second to give a practical way to estimate such shapes in real time. Experiments on various sets of data including human body silhouettes are shown to illustrate the principle and the interests of visual shapes.