Geometric representations and deep Gaussian conditional random field networks for computer vision