Official PyTorch implementation of our work: “Continual coarse-to-fine domain adaptation in semantic segmentation”, published in Elsevier Image and Vision Computing, Volume 121, May 2022.
Corresponding author: donald.shenaj@dei.unipd.it
This repository uses the following libraries:
- Python (3.7.11) [tested up to 3.9.7]
- PyTorch (1.10.0) [tested up to 1.10.1]
- torchvision (0.11.1) [tested up to 0.11.2]
- tensorboardX (2.4) [tested up to 2.4.1]
- opencv (4.5.4-dev) [tested up to 4.5.5]
- numpy (1.21.2)
- pillow (8.4.0)
- tqdm (4.62.3)
All the dependencies are listed in the requirements.txt. To install the requirements:
pip install -r requirements.txtDownload the GTA5 dataset and place it inside the folder ${dataset_root}/GTA. The dataset split is provided within the train.txt and val.txt files in the folder datasets/splits/GTA.
Download the Cityscapes dataset (fine and coarse) and place it inside the folder ${dataset_root}/Cityscapes. The dataset split is provided within the train.txt, val.txt and test.txt files in the folder datasets/splits/Cityscapes.
Download the IDD dataset and place it inside the folder${dataset_root}/IDD. The dataset split is provided within the train.txt, val.txt and test.txt files in the folder datasets/splits/IDD.
Download the official PyTorch ResNet101's weights pretrained on ImageNet from here and place them inside the folder models/backbone_checkpoints.
- Source Only
python -u train.py --logdir "log/c2f/gta_to_cityscapes/source_only" --incremental_setup "c2f" --classifier "DeepLabV3" --blur_mul 5 --backbone "ResNet101" --incremental_iterations 25000 --dataloader_workers 4 --c2f_weight_init False --kd_lambda 0 --kd_lambda_c2f 0 --msiw_lambda 0 --validate_every_steps 2500 --batch_size 2 --source_dataset "gta" --target_dataset "cityscapes" --source_root_path "${dataset_root}/GTA" --source_splits_path "datasets/splits/GTA" --target_root_path "${dataset_root}/Cityscapes" --target_splits_path "datasets/splits/Cityscapes" - SKDC
python -u train.py --logdir "log/c2f/gta_to_cityscapes/skdc" --incremental_setup "c2f" --classifier "DeepLabV3" --blur_mul 5 --backbone "ResNet101" --incremental_iterations 25000 --dataloader_workers 4 --c2f_weight_init True --kd_lambda 10 --kd_lambda_c2f 10 --msiw_lambda 0 --kd_type "kdc" --validate_every_steps 2500 --batch_size 2 --source_dataset "gta" --target_dataset "cityscapes" --source_root_path "${dataset_root}/GTA" --source_splits_path "datasets/splits/GTA" --target_root_path "${dataset_root}/Cityscapes" --target_splits_path "datasets/splits/Cityscapes" - CCDA
python -u train.py --logdir "log/c2f/gta_to_cityscapes/ccda" --incremental_setup "c2f" --classifier "DeepLabV3" --blur_mul 5 --backbone "ResNet101" --incremental_iterations 25000 --dataloader_workers 4 --c2f_weight_init True --kd_lambda 10 --kd_lambda_c2f 10 --msiw_lambda 0.1 --uda_skip_iterations 100 --kd_type "kdc" --validate_every_steps 2500 --batch_size 2 --source_dataset "gta" --target_dataset "cityscapes" --source_root_path "${dataset_root}/GTA" --source_splits_path "datasets/splits/GTA" --target_root_path "${dataset_root}/Cityscapes" --target_splits_path "datasets/splits/Cityscapes" - Source Only
python -u train.py --logdir "log/c2f/gta_to_IDD/source_only" --incremental_setup "c2f" --classifier "DeepLabV3" --blur_mul 5 --backbone "ResNet101" --incremental_iterations 25000 --dataloader_workers 4 --c2f_weight_init False --kd_lambda 0 --kd_lambda_c2f 0 --msiw_lambda 0 --validate_every_steps 2500 --batch_size 1 --source_dataset "gta" --target_dataset "IDD" --source_root_path "${dataset_root}/GTA" --source_splits_path "datasets/splits/GTA" --source_rescale_size "1280,720" --target_rescale_size "1280,720" --target_root_path "${dataset_root}/IDD" --target_splits_path "datasets/splits/IDD" - SKDC
python -u train.py --logdir "log/c2f/gta_to_IDD/skdc" --incremental_setup "c2f" --classifier "DeepLabV3" --blur_mul 5 --backbone "ResNet101" --incremental_iterations 25000 --dataloader_workers 4 --c2f_weight_init True --kd_lambda 10 --kd_lambda_c2f 1 --msiw_lambda 0 --kd_type "kdc" --validate_every_steps 2500 --batch_size 1 --source_dataset "gta" --target_dataset "IDD" --source_root_path "${dataset_root}/GTA" --source_splits_path "datasets/splits/GTA" --source_rescale_size "1280,720" --target_rescale_size "1280,720" --target_root_path "${dataset_root}/IDD" --target_splits_path "datasets/splits/IDD" - CCDA
python -u train.py --logdir "log/c2f/gta_to_IDD/ccda" --incremental_setup "c2f" --classifier "DeepLabV3" --blur_mul 5 --backbone "ResNet101" --incremental_iterations 25000 --dataloader_workers 4 --c2f_weight_init True --kd_lambda 10 --kd_lambda_c2f 1 --msiw_lambda 0.1 --uda_skip_iterations 100 --kd_type "kdc" --validate_every_steps 2500 --batch_size 1 --source_dataset "gta" --target_dataset "IDD" --source_root_path "${dataset_root}/GTA" --source_splits_path "datasets/splits/GTA" --source_rescale_size "1280,720" --target_rescale_size "1280,720" --target_root_path "${dataset_root}/IDD" --target_splits_path "datasets/splits/IDD" python -u test.py --logdir "$train_path/test" --step_model_checkpoint "${train_path}/ResNet101_DeepLabV3_step0.pth" --incremental_setup "c2f" --classifier "DeepLabV3" --backbone "ResNet101" --batch_size 1 --target_val_split "test" --target_dataset "cityscapes" --target_root_path "${dataset_root}/Cityscapes" --target_splits_path "datasets/splits/Cityscapes" python -u test.py --logdir "$train_path/test" --step_model_checkpoint "${train_path}/ResNet101_DeepLabV3_step0.pth" --incremental_setup "c2f" --classifier "DeepLabV3" --backbone "ResNet101" --batch_size 1 --target_val_split "test" --target_rescale_size "1280,720" --target_dataset "IDD" --target_root_path "${dataset_root}/IDD" --target_splits_path "datasets/IDD/splits" If you find our work relevant to your research and you use this repository, please consider to cite:
@article{shenaj2022continual,
title = {Continual coarse-to-fine domain adaptation in semantic segmentation},
author = {Donald Shenaj and Francesco Barbato and Umberto Michieli and Pietro Zanuttigh},
journal = {Image and Vision Computing},
volume = {121},
pages = {104426},
year = {2022}
}