Volumetric reconstruction by random sampling

The problem addressed in this paper concerns the reconstruction from a limited number of views of a 3D scene composed of a finite number of differentiable manifolds. Each manifold is considered as a realization of a Markov random field and the reconstruction is based on sampling the a-priori density of the field. Measurement are assumed to be the photo consistency of small patches of the tangent bundle to the manifold back projected on the cameras. The samples that maximize the a-posteriori density of the field are kept and used to update the visibility maps and the reconstruction. Some examples of surface reconstructions by the application of the proposed method are shown.