This document discusses semantic segmentation using weak supervision from bounding boxes. It presents an approach called "Simple Does It" that trains a segmentation network on bounding box annotations instead of full pixel-level labels. The network learns to generate segmentation masks and classify object categories within bounding boxes. Related work on weakly supervised segmentation is also mentioned. Evaluation shows the approach achieves 20% accuracy compared to human labels, and combining with GrabCut post-processing improves results further.