By shape we mean the surface of objects when they are observed from several viewpoints. Again depending on the primitives considered, shape will be partially or completely described. We will particularly focus on silhouettes for several reasons: They lead directly to 3D shapes and not 3D points as it has extensively been done in the recent past; They represent a robust source of information, an object always projects onto a silhouette whatever its appearance is, and partial occlusions do not impede 3D estimations; An explicit matching between viewpoints is not required for 3D computations; Coherence principles explained before should improve drastically their extractions in 2D images.

Classically, silhouettes are used to compute visual hulls which are object shape approximations. This process is well known and we have recently proposed an exact solution for that. However, the visual hull is not the object shape but its approximation, and a scientific challenge, still to be tackled, is the recovery of exact object shapes. This require more information than the one provided by silhouettes which are already entirely encoded into the visual hull. Such additional pieces of information are, for instance, shape smoothness information or concavity information. We plan to investigate this direction, in particular through promising geometric tools such as subdivision surfaces.