Some Random Thoughts

[michaelku1]

I personally get your frustration as I think the cityscapes to foggy cityscapes adaptation is hard to achieve good semantic alignment on due to the amount of noise that distorts the semantic features.

[wymanCV]

Hi, thanks for your interest.
I think the fluctuation is more likely caused by the EMP baseline since the multi-level adversarial learning seems to have a more obvious influence on the pixel-level object detector. We found that the phenomenon can be relieved from four aspects: reducing the learning rate, training enough iterations, using Style-GAN-generated data, and using all-level Foggy Cityscapes data instead of only 0.02.

This phenomenon is also common in DA-FRCNN, even for the single-class KITTI scene (here). Hence, I really hope to see new DAOD works can solve this issue, since this also gives me a headache.

[michaelku1]

I see. So I guess it's just minor tuning for the time being. Thanks for your insights. I agree that there is still some work to do in order to succeed in adversarial training over semantic alignment.