Inconsistency between the paper and the implementation

[naoto0804]

I'm trying to re-implement your model on PyTorch and I have a few questions about details.

• Feature aggregation
  In Fig. 2, netaggi_0 / netaggm_0 seems to be concatenated. In the code, it is not concatenated. Which should I follow?

• Missing 1x1 conv?
  In Fig. 2, 1x1 conv. layer is inserted before SPP module. In the code, it's not. Should I add it?

[vinthony]

Thanks for your useful help.

You can follow the model from the code. The results are based on the model in this repo.

It seems that there are some small mistakes in the figure although it might not highly influence the performance.

I will try to retrain the model based on the model from paper.
and add some Introduction in the README about these inconsistencies.

Thanks a lot!

[naoto0804]

Sorry to ask you one more;

Can you elaborate a bit more about why you divide the losses by some magic numbers?
(here)

[vinthony]

These numbers are following the setting of the framework of our baseline code in reflection removal here

[vinthony]

@naoto0804
I re-train the model with the model in the paper (Fig.2) on the SRD dataset without data augmentation for a quick experiment.
The performance improves a bit on the Non-Shadowed Region (NS in the following Table), while the performance in the shadowed region is the same. However, the setting from the paper uses more parameters. Here is the RMSE report:

Method	S	NS	ALL
Figure.2	8.94	4.40	5.35
This repo	8.94	4.80	5.67

So I suggest you use the model from the paper.

[naoto0804]

Wow, thank you so much!

[naoto0804]

Why do you initialize almost all the convolutional layers by identity_initializer?

[vinthony]

Other initializers might also work.
We follow the baseline of CAN in fast image processing: https://github.com/CQFIO/FastImageProcessing

[naoto0804]

Is it possible to share the masks that you used for the SRD dataset, which is essential for evaluation? I'm trying to extract it, but my result looks too noisy to be a ground truth;

[vinthony]

Hi, we also use a simple rule to extract the mask from the image pair.
Although it is not totally correct, we found it still helps for shadow removal.
Here is the link of the mask of SRD dataset. OneDrive

[naoto0804]

I really really appreciate it;