Awesome

ASSUDA

Code and data of Exploring Robustness of Unsupervised Domain Adaptation in Semantic Segmentation (ICCV 2021; Oral)

Datasets

• GTA5 -> Cityscapes (VGG) and GTA5 -> Cityscapes (DeepLab)
• SYNTHIA -> Cityscapes (VGG) and SYNTHIA -> Cityscapes (DeepLab)
• Cityscapes
• Perturbed test images: PSPNet_Attack

Initial models

• VGG
• DeepLab

Training

An example (SYNTHIA->Cityscapes with DeepLab):

python main.py \
    --data-dir /path/to/synthia_deeplab \
    --data-list ./dataset/synthia_list/train.txt \
    --data-dir-target /path/to/cityscapes \
    --data-list-target ./dataset/cityscapes_list/train.txt \
    --data-label-folder-target /path/to/synthia_deeplab/cityscapes_ssl \
    --snapshot-dir ./snapshots/synthia2city_deeplab \
    --init-weights ./initial_model/DeepLab_init.pth \
    --num-steps-stop 80000 \
    --model DeepLab \
    --source synthia \
    --learning-rate 1e-4 \
    --learning-rate-D 1e-6 \
    --lambda-adv-target 1e-4 \
    --save-pred-every 5000 \
    --alpha 1.0 \
    --lambda-contrastive 0.01

Evaluation

An example (SYNTHIA->Cityscapes with DeepLab):

python evaluation.py \
    --data-dir-target /path/to/pspnet_attack/pspnet_fgsm_0.1 \
    --data-list-target ./dataset/cityscapes_list/val.txt \
    --gt_dir /path/to/cityscapes/gtFine/val \
    --devkit_dir ./dataset/cityscapes_list \
    --restore-from ./snapshots/synthia2city_deeplab/synthia_80000 \
    --save results/cityscapes_eval \
    --model DeepLab \
    --source synthia

Citation

@article{yang2021exploring,
  title={Exploring Robustness of Unsupervised Domain Adaptation in Semantic Segmentation},
  author={Yang, Jinyu and Li, Chunyuan and An, Weizhi and Ma, Hehuan and Guo, Yuzhi and Rong, Yu and Zhao, Peilin and Huang, Junzhou},
  journal={Proceedings of the IEEE international conference on computer vision (ICCV)},
  year={2021}
}

Acknowledgment

The code is heavily borrowed from BDL