Multi-view stereo is the problem of recovering the shape of objects from multiple images taken from different but known camera positions. It is an inverse problem where we want to find the cause (the object) given the effect (the images). From a Bayesian perspective, the solution would be the reconstruction that best reproduces the input images while at the same time being plausible a priori. Taking this approach, in this thesis we develop generative models and methods for computing reconstructions that minimize the difference between the observed images and the images sythetized from the reconstruction.

Three models are presented. The first, represents the reconstructed scene by a set of depth maps. This gives high resolution results, but have problems at the objects boundaries. The second model represents the scene by a discreet occupancy grid, yielding to a combinatorial optimization problem, which is addressed through message passing techniques. The final model represents the scene by a smooth surface and the resulting optimization problem is solved via gradient descent surface evolution.

In either model, the main difficulty is to correctly take into account the occlusions. Modeling self-occlusions results in optimization problems that challenge current optimization techniques. In this respect, the main result of the thesis is the computation of the derivative of the reprojection error with respect to surface variations taking into account the visibility changes that occur while the surface moves. This enables the use of gradient descent techniques, and leads to surface evolutions that place the contour generators of the surface to their correct location in the images without the need of additional silhouettes or apparent contours constraints.