About Eq.(2) proposed in the paper...

[Circle0905]

Hello,

This is an interesting paper, but I encountered some issues while studying it.

In the paper, the context affinity loss (L_{ca}) encourages visually dissimilar pixels to have different labels, and vice versa, as stated in Equations (2) and (3).

However, in the code (specifically in feature_loss.py and train_processes.py), the predictions obtained from the model are used without any additional operations. How is the operation proposed in Equation (2) represented in the code?

Thank you!

[dddraxxx]

It is somewhere in the code. But not the same name as in the paper. I think you can pay attention to here, which give the args of the visual kernel.

[Circle0905]

Could you please clarify how Equ. (2) is implemented in the code?

I reviewed the featureloss.py file and noticed that D(i, j) is defined as p{i} - p_{j}, which appears to be equivalent to the Local Saliency Coherence Loss as introduced in the SCWSSOD (Structure-Consistent Weakly Supervised Salient Object Detection with Local Saliency Coherence). However, this seems to differ from what is described in Equ. (2). I am concerned that there might be an oversight or a misunderstanding.

Thank you!

[dddraxxx]

1. The loss actually originates from Gated CRF Loss for Weakly Supervised Semantic Image Segmentation, and is widely adopted by numerous weakly supervised papers.
2. I don't know what you are referring. It will be easier for me to check if you could include the link to code in your question. Thanks.