search

uta-smile / ASSUDA

Exploring Robustness of Unsupervised Domain Adaptation in Semantic Segmentation (ICCV 2021; Oral)
MIT License
12 stars 1 forks source link

Question2 #5

Open JayShao-Xie opened 2 years ago

JayShao-Xie commented 2 years ago

Excuse me.

  1. Dataset GTA5 -> Cityscapes (DeepLab) doesn't contain 'synthia_mapped_to_cityscapes' and 'cityscapes_ssl' folder.
  2. As mentioned in paper, we dont need target label, but in main.py Line 79"trg_img_adv = fgsm.untargeted(model, trg_img, trg_lbl)" we need target label for training. Is this target label pseudo labels? If I use another target dataset without labels for training, how can I deal with this problem.
viyjy commented 2 years ago

Sorry for the late reply.

  1. Since it is GTA5 to Cityscapes, it does not contain synthia.
  2. Yes, it is pseudo labels. You can first train an UDA model, then use the trained model to generate pseudo labels. Feel free to let me know if you might need additional information. Thanks.