search

Haochen-Wang409 / U2PL

[CVPR'22] Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels
Apache License 2.0
425 stars 59 forks source link

For domain adaptation #144

Closed tanveer6715 closed 1 year ago

tanveer6715 commented 1 year ago

Hi Thank you for sharing this great work. I just want to know have you tried this method for domain adaptation? Can we used it? If yes then what will be the changes required according to your suggestions? Thanks

Haochen-Wang409 commented 1 year ago

I think the key challenge of domain adaptation is not the same as semi-supervised learning. Therefore, it seems that U2PL can be used in DA but I am not sure about the performance.

For DA, just modify the labeled dataset and the unlabeled dataset.

tanveer6715 commented 1 year ago

Thanks for your response. In U2PL I guess the unlabeled data is cutmixed. However for domain adaptation can we cutmixed labeled data(source domain) into unlabeled data (target domain)?

Haochen-Wang409 commented 1 year ago

Yes, of course we can. [1] even applied ClassMix between source images and target images.

[1] DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation. In CVPR 2022.

tanveer6715 commented 1 year ago

Thank you. I am closing this issue. :)