Bidirectional Self-Training with Multiple Anisotropic Prototypes for Domain Adaptive Semantic Segmentation

(ACM MM 2022) This is a pytorch implementation of [BiSMAP](Bidirectional Self-Training with Multiple Anisotropic Prototypes for Domain Adaptive Semantic Segmentation).

Prerequisites

• Python 3.7.0
• GPU Memory >= 11G (Preferably using a gpu with tensor core)
• Pytorch 1.8.2

Getting started

• Download The GTA5 Dataset

• Download The SYNTHIA Dataset

• Download The Stylized Datasets

• Download The Cityscapes Dataset

• Download The pretrained model

The data folder is structured as follows:

├── dataset/
│   ├── CityScape/     
|   |   ├── gtFine/
|   |   ├── leftImg8bit/
│   ├── GTA5/
|   |   ├── images/
|   |   ├── labels/
│   ├── gta_stylized/
│   └── 			
└── pretrained/
│   ├── r101_1x_sk0.pth
│   ├── r152_1x_sk1.pth
│   ├── gta5_dill_model.pkl
│   ├── gta5_warmup_model.pkl
...

Evaluate models

Our final model is available via MEGA

python3 test.py --config ./configs/gta5_test.yml 

---

Train models

Generate Source Transferability Map and Pseudo Labels

python source_transferability_map.py --config ./configs/gta5_stm.yml

python generate_maps_pseudolabel.py --config configs/gta5_maps_pla.yml --threshold 100

Self-Training

python train.py --config configs/gta5_st.yml --logdir ./runs/gta5_st

Self-Distillation (two stage)

# distill 1
python3 inference.py --config ./configs/inference.yml --checkpoint_path ./runs/gta5_st/from_gta5_to_cityscapes_on_deeplabv3_plus_best_model.pkl --save_path pseudolabels_dill
python3 train.py --gpu 1 --config ./configs/gta5_dill.yml --dill_teacher ./runs/gta5_st/from_gta5_to_cityscapes_on_deeplabv3_plus_best_model.pkl --logdir ./runs/gta5_dill_1

# distill 2
python3 inference.py --config ./configs/inference.yml --checkpoint_path ./runs/gta5_dill_1/from_gta5_to_cityscapes_on_deeplabv3_plus_best_model.pkl --save_path pseudolabels_dill
python3 train.py --gpu 1 --config ./configs/gta5_dill.yml --dill_teacher ./runs/gta5_dill_1/from_gta5_to_cityscapes_on_deeplabv3_plus_best_model.pkl --logdir ./runs/gta5_dill_2

Citation

If you use this code in your research please consider citing

@article{lu2022bidirectional,
  title={Bidirectional Self-Training with Multiple Anisotropic Prototypes for Domain Adaptive Semantic Segmentation},
  author={Lu, Yulei and Luo, Yawei and Zhang, Li and Li, Zheyang and Yang, Yi and Xiao, Jun},
  journal={arXiv preprint arXiv:2204.07730},
  year={2022}
}