Comparison of the of FCN network with different skip architectures, based on the pre-trained VGG 16-layer net.

Learning is end-to-end, pixel-to-pixel.