Table 3.

Architecture of our networks. “Stride” is the rate of up/down sampling. When “Stride” is $2\times 2$, the size of the activation map is reduced by half. “fc” in Type indicates a fully connected layer. In the “conv” type of (b), the “Kernel” is $2\times 2$ and the “Stride” is $1\times 1$. In the “maxpooling” type of (b), the “Kernel” is $2\times 2$ and the “Stride” is $2\times 2$.

(a) Architecture of ICNET					(b) Architecture of CDNET
Type	Kernel	Dilation Rate $\mathbf{k}$	Stride	${\mathrm{Channel}}_{\mathrm{out}}$	Type	${\mathrm{Channel}}_{\mathrm{out}}$
conv	$3\times 3$	1	$2\times 2$	32	conv+maxpooling	64
conv	$3\times 3$	1	$2\times 2$	64	conv+maxpooling	128
conv	$3\times 3$	1	$1\times 1$	128	conv+maxpooling	256
dilated conv	$3\times 3$	2	$1\times 1$	128	fc	512
dilated conv	$3\times 3$	4	$1\times 1$	128	fc	256
conv	$3\times 3$	1	$1\times 1$	256	fc	128
conv	$3\times 3$	1	$1\times 1$	256	fc	1
transposed conv	$1\times 1$	1	$1/2\times 1/2$	128
conv	$3\times 3$	1	$1\times 1$	64
conv	$3\times 3$	1	$1\times 1$	32
conv	$3\times 3$	1	$1\times 1$	3