Can you share the implementation of the compared SemiCDNet method？

wgcban / SemiCD

Revisiting Consistency Regularization for Semi-supervised Change Detection in Remote Sensing Images

https://www.wgcban.com/research#h.ar24vwqlm021

125 stars 19 forks source link

Can you share the implementation of the compared SemiCDNet method？ #2

Closed lauraset closed 2 years ago

lauraset commented 2 years ago

Hi, @wgcban . I learned a lot from your work. I want to know whether you can share the implementation of the SemiCDNet method you compared? Thank you very much.

wgcban commented 2 years ago

Hi @lauraset

Thank you for your interest in our work, and I am happy that you learned from our work.

The state-of-the-art methods that I have used for the comparisons in my paper are not publically available. Hence, I have implemented those methods on my own and you can find all of those methods by shifting the branch.

s4GAN: change the branch from main to s4GAN: https://github.com/wgcban/SemiCD/tree/s4GAN
SemiCDNet (TGRS-21): change the branch from main to SemiCDNet(TGRS21): https://github.com/wgcban/SemiCD/tree/SemiCDNet(TGRS21)
AdvNet: Change the branch from main to AdvNet: https://github.com/wgcban/SemiCD/tree/AdvNet
RoCD: Change the branch from main to RoCD: https://github.com/wgcban/SemiCD/tree/RoCD

I hope this answers your question. Please let me know if you have more questions!

Best, Chaminda.

lauraset commented 2 years ago

Hi, @wgcban . Thank you for your detailed response. You are really a nice person. One issue confused me. That is the high similarity between your method with the CCT (cross-consistency training) method. Actually, I guess that the temporal consistency (e.g., the high similarity of unchanged areas and the high dissimilarity of changed areas between bi-temporal images) embedded in the change detection is not utilized yet. I guess this point may differentiate the change detection task from the semantic segmentation one for the CCT method. Anyway, thank you for your help.

Best regards, Yinxia Cao.

wgcban commented 2 years ago

@lauraset

Yes, you are correct that the proposed method is somewhat similar to CCT; however, the main idea that we want to convey from this work is that the cluster assumption exists even for bitemporal images in the deep feature difference domain and consistency training can be used for semi-supervised (SS) change detection, while CCT considers SS semantic segmentation with RGB images, not bi-temporal images.

Yes, I agree with your point that our method does not integrate the time consistency of change and no-change regions, and I will address this in my future work. Thanks for your suggestion.

Best, Chaminda.

lauraset commented 2 years ago

@wgcban , Thank you for your nice help. I have no further questions.

wgcban commented 2 years ago

Nice, I am closing this issue.