What's Behind the Mask: Estimating Uncertainty in Image-to-Image Problems