What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?