A convex multi-view stereo formulation with robustness to occlusions and time-varying clutter

We present a novel framework for multi-view stereo that poses the problem of recovering a 3D surface in the scene as a regularized minimal partition problem of the visibility function in the presence of clutter. We introduce a simple and robust method to integrate estimates from several views that tolerates both static and time-varying clutter. Our formulation does not rely on the visual hull, 2D silhouettes, or make use of initial surface estimates. Furthermore, we use a globally optimal framework, so that the solution does not depend on initialization and computationally efficient numerical methods can be used to find the solution. We also strive for simplicity so that more general models of image formation can be used without compromising the estimation process. Experimental results on synthetic and publicly available real data show that our method performs on a par with state-of-the-art methods that have been used on clutter-free data. © 2010 ACM.