论文中的一些疑惑

[wuhuikai]

论文中的一些疑惑

[wuhuikai]

1. y_s 是没有dilated convolution的backbone的输出, 即经典FCN的backbone的输出
2. h是通过一个CNN (即JPU) 建模的。具体来说，JPU是在进行如下操作: (1)求解出一个隐式的h, (2)使用隐式的h来处理y_m, 获得最后上采样后的特征图

[HolmesShuan]

Hi, SOTA with 3x speedup. Nice job!
Due to my limited knowledge of segmentation, I find the following descriptions hard to understand, any help?

1. The caption of Figure 5: ... and the rest part of y_m. What do you mean the rest part?
2. The right subfigure of Figure 5 wraps node 1,2,3 with 3x3 conv, d=1, yet left subfigure uses node 0,2,4. Note that dilation rate 1 focuses on y^0_m, which corresponds to the left subfigure or the other? How to understand dilation rate 1 focuses on y^0_m and d=2 aims at y^0_m and y_s?
3. The problem defined in Eq.(1): is a known function? If it is true, Eq.(4) with learnable and may be different from the pre-defined joint upsampling problem.
4. As described in section 3.3.3, Eq.(4) is an optimization problem. What is the specific learning target in this problem?

Thanks in advance.

[wuhuikai]

1. As shown in Equation 4, y_m can be split into y_m^0 and y_m^1. The rest part of y_m means y_m^1.
2. Both d=1 and d=2 correspond to the right figure. The left figure shows that y_s only depends on y_m^0. The inputs of d=1 (the dotted box) are y_m^0 and y_m^1, while the inputs of d=2 (the solid box) are y_m^0.
3. f() is a know function. As for Equation 4, y_s = h(y_m^0) is also a know function after the training converges.
4. The objective is to find a function \hat{h}() to approximate y_s = h(y_m^0).

[HolmesShuan]

Thanks for your responses. I still wonder the proposed solving with CNNs indeed solves the optimization problem strictly? The stage (b) seems to optimize and simultaneously.

[wuhuikai]

JPU does not solve the optimization problem strictly. Motivated by joint upsampling, we employ a CNN to approximate the optimization problem.

[HolmesShuan]

Cool, thanks.