feature space adaptation

[jianingwangind]

Hi, thanks for the great work and the codes. In the original implementation, i only found the output space adaptation, since the both outputs from deeplab_multi are coming from the classifier layers and the upsampled for the later alignment. So if i want to implement the real feature space adaptation, i should take the feature maps right before the classifier layers, right? And also a question, the output of from layer5, in my opinion, is not shallow enough, would aligning the features coming from very shallow layers help the domain adaptation?

looking forward to your reply.

[wasidennis]

@jianingwangind yes, if you want to do feature-level adaptation, you should use features before the classifier.

For shallow layers, we have not tried to use them, but in our multi-level model, it also considers a lower-level output and it improves results a bit.

[jianingwangind]

@wasidennis thanks a lot for your fast reply:) May i ask you, in your opinion, how would feature space adaptaion and output space adaptation affect the final performance differently?

[wasidennis]

We have shown in the paper that, feature-level adaptation could be less stable during adversarial training. In terms of performance, we also show in the paper that output space adaptation performs better.

[jianingwangind]

@wasidennis thanks again for the great reply. Now i get it.