about the size of predict_result

[jjkkqq]

hi! from your paper, i see the final size of your predict is 1/2 of the original size, is it necessary to conduct an upsampling op? thank you!

[AndreKelm]

Hello, yes, upsampling is necessary, but it is already present in the implementation and used in the code. I think it was the bicubic one.

[jjkkqq]

I am studying the impact of upsampling on edge detection. Thank you for your reply！

[CoachingJane]

Due to equipment problems, I can't see your code. From your structure diagram, is thr Conv module just convolution instead of convolution plus activation function (such as relu)? Because you only marked the relu function on the RCU module. thankyou！

[AndreKelm]

It is just convolution! To understand the idea behind the different blocks and activation layers, I copy the corresponding text section from the RefineNet Paper: „Note that all convolutional components of the RefineNet have been carefully constructed inspired by the idea behind residual connections and follow the rule of identity map- ping [25]. This enables effective backward propagation of the gradient through RefineNet and facilitates end-to-end learning of cascaded multi-path refinement networks. Employing residual connections with identity mappings allows the gradient to be directly propagated from one block to any other blocks, as was recently shown by He et al. [25]. This concept encourages to maintain a clean information path for shortcut connections, so that these connections are not “blocked” by any non-linear layers or components. In- stead, non-linear operations are placed on branches of the main information path. We follow this guideline for de- veloping the individual components in RefineNet, including all convolution units. It is this particular strategy that allows the multi-cascaded RefineNet to be trained effectively. Note that we include one non-linear activation layer (ReLU) in the chained residual pooling block. We observed that this ReLU is important for the effectiveness of subsequent pool- ing operations and it also makes the model less sensitive to changes in the learning rate. We observed that one single ReLU in each RefineNet block does not noticeably reduce the effectiveness of gradient flow.“ [Lin, G., Milan, A., Shen, C., Reid, I.: Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5168–5177 (July 2017). https://doi.org/10.1109/CVPR.2017.549]