Justin Theiss, Jay Leverett, Daeil Kim, Aayush Prakash
In ECCV 2022 (Oral).

Source GTA5	GTA5 Translated with VSAIT

Clone this repo:

git clone https://github.com/facebookresearch/vsait.git
cd vsait/

Install dependencies via pip:

pip install -r requirements.txt

For any two image datasets with png/jpg images, download source and target data (or create symlinks) to ./data/source/ and ./data/target/ with train and val subfolders for each domain.

For gta2cityscapes, GTA5 dataset images folder should be split into training and validation folders to be stored in ./data/source/train/ and ./data/source/val/, respectively. Similarly, the Cityscapes dataset folders /leftImg8bit/train/ and /leftImg8bit/val/ should be stored in ./data/target/train/ and ./data/target/val/, respectively.

Launch training with defaults in configs:

python train.py --name="vsait"

This will use the default configs in ./configs/ and save checkpoints and translated images in ./checkpoints/vsait/.

Translate images in ./data/source/val/ using a specific checkpoint:

python test.py --name="vsait_adapt" --checkpoint="./checkpoints/vsait/version_0/checkpoints/epoch={i}-step={j}.ckpt"

Images from the above example would be saved in ./checkpoints/vsait_adapt/images/.

VSAIT is released under the CC-BY-NC 4.0 License.