Performance on the burrow dataset. We picture one 2D slice of the original volume (top left), as well as human (top right) and machine annotated segmentation masks (bottom right). On the bottom left, we render the residual errors.