Table 2. 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.

Evaluation metrics	FCN-32s	FCN-16s	FCN-8s
pixel acc.	0.923	0.928	0.935
mean acc.	0.795	0.811	0.807
mean IU	0.728	0.749	0.752
f.w. IU	0.876	0.886	0.892
inference time	0.318s	0.317s	0.323s