Paper #5: The Front End and the Context Module
The authors removed VGG-16's last two shrinking steps and beat earlier segmentation networks by more than 5 points. Then they added a small stack of dilated layers, started so each one passes its input through unchanged, and gained more.