On Regularized Losses for Weakly-supervised CNN Segmentation