What is the difference between DODA(Only VSS) and DODA (w/o TACM )

[xiaodongww]

Hi, thanks for your work. I have a small question on Table 1 of your paper.

You reported three results DODA (only VSS), DODA(w/oTACM), DODA. From my understanding, there are two modules VSS and TACM, then what is the difference between DODA(Only VSS) and DODA (w/o TACM )? Do I miss something else?

Thanks!

[Dingry]

Hi, DODA (only VSS) means the model is pretrained with VSS without self-training. DODA (w/o TACM) means the model is pretrained with VSS and self-trained without TACM.

[xiaodongww]

Thanks for your quick reply! @Dingry
Here is another question. Is the self-training strategy orthogonal to existing UDA methods like MCD or MMD? Will the performance of DODA be further improved if we apply MMD loss on backbone feature learning?

[Dingry]

Hi, I think they are orthogonal, since TACM is applied at the data level. But I haven't tried MCD or MMD with TACM. You can try it if you are interested and looking forward to your results!

[xiaodongww]

Thanks for your reply very much. Do you have a plan for when to release the code of the baseline UDA methods? I tried to implement MCD for domain adaptation by myself but found the training process corrupted.

[Dingry]

Hi, thanks for your attention. We have plan to release baseline methods but don't have exact time now since we are working with another paper. We will try to re-organize code and release it later. Thanks!