In this work, we propose ATP, a novel source data-free adaptation framework for semantic segmentation tasks. The proposed method consists three steps: curriculum feature alignment, complementary self-training, and information propogation. Extensive experiments have demonstrate that the prpoposed ATP boosts the performance of source data-free domain adaptation tasks.

Following the repo of SePiCo, the environment is conducted as following.

This code is implemented with Python 3.8.5 and PyTorch 1.7.1 on CUDA 11.0.

To try out this project, it is recommended to set up a virtual environment first:

# create and activate the environment
conda create --name ATP -y python=3.8.5
conda activate ATP

# install the right pip and dependencies for the fresh python
conda install -y ipython pip

Then, the dependencies can be installed by:

# install required packages
pip install -r requirements.txt

# install mmcv-full, this command compiles mmcv locally and may take some time
pip install mmcv-full==1.3.7  # requires other packeges to be installed first

Alternatively, the mmcv-full package can be installed faster with official pre-built packages, for instance:

# another way to install mmcv-full, faster
pip install mmcv-full==1.3.7 -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7.0/index.html

The environment is now fully prepared.

• GTAV: Download all zipped images, along with their zipped labels, from here and extract them to a custom directory.
• Cityscapes: Download leftImg8bit_trainvaltest.zip and gtFine_trainvaltest.zip from here and extract them to a custom directory.
• ACDC: Download rgb_anon_trainvaltest.zip and gt_trainval.zip from here and extract them to a custom directory.

Symlink the required datasets:

ln -s /path/to/gta5/dataset data/gta
ln -s /path/to/cityscapes/dataset data/cityscapes

Perform preprocessing to convert label IDs to the train IDs and gather dataset statistics:

python tools/convert_datasets/gta.py data/gta --nproc 8
python tools/convert_datasets/cityscapes.py data/cityscapes --nproc 8

Ultimately, the data structure should look like this:

ATP
├── ...
├── data
│   ├── cityscapes
│   │   ├── gtFine
│   │   ├── leftImg8bit
│   ├── dark_zurich
│   │   ├── corresp
│   │   ├── gt
│   │   ├── rgb_anon
│   ├── gta
│   │   ├── images
│   │   ├── labels
├── ...

Below, we provide checkpoints of ATP for different benchmarks.

To begin with, download SegFormer's official MiT-B5 weights (i.e., mit_b5.pth) pretrained on ImageNet-1k from here and put it into a new folder ./pretrained.

The training entrance is at run_experiments.py. To examine the setting for a specific task, please take a look at experiments.py for more details. Generally, the training script is given as:

python run_experiments.py --exp <exp_id>

To evaluate the pretrained models on Cityscapes, please run as follows:

python -m tools.test /path/to/config /path/to/checkpoint --eval mIoU

This project is based on the following open-source projects. We thank their authors for making the source code publicly available.

• SePiCo
• MMSegmentation
• SegFormer
• DAFormer
• DACS
• DANNet

For help and issues associated with ATP, or reporting a bug, please open a [GitHub Issues], or feel free to contact [email protected].