chjXu/CoDA

CoDA: Instructive Chain-of-Domain Adaptation with Severity-Aware Visual Prompt Tuning

🌟🌟🌟 Here is the official project of 🎻CoDA. We only release the checkpoint for inference now and will release the code of Chain-of-Domain and Severity-Aware Visual Prompt Tuning later.

🔥🔥🔥CoDA is a UDA methodology that boosts models to understand all adverse scenes (☁️,☔,❄️,🌙) by highlighting the discrepancies within these scenes. CoDA achieves state-of-the-art performances on widely used benchmarks.

Experiments	mIoU	Checkpoint	Configs
Cityscapes $\rightarrow$ ACDC	72.6	-	-
Cityscapes $\rightarrow$ Foggy Zurich	60.9	-	-
Cityscapes $\rightarrow$ Foggy Driving	61.0	-	-
Cityscapes $\rightarrow$ Dark Zurich	61.2	-	-
Cityscapes $\rightarrow$ Nighttime Driving	59.2	-	-
Cityscapes $\rightarrow$ BDD100K-Night	41.6	-	-

Download Checkpoint

cd CoDA
python ./tools/download_ck.py

or you can manually download checkpoints from Google Drive.

Environment

Before run demo, first configure the PYTHONPATH, or you will encounter error like 'can not found tools...'.

cd CoDA
export PYTHONPATH=.:$PYTHONPATH

or directly modify the .bashrc file

vi ~/.bashrc
export PYTHONPATH=your path/CoDA:$PYTHONPATH
source ~/.bashrc

demo

python ./tools/image_demo.py --img ./images/night_demo.png --config ./configs/coda/csHR2acdcHR_coda.py --checkpoint ./pretrained/CoDA_cs2acdc.pth

Inference Steps

python ./tools/image_demo.py --img_dir ./acdc_dir --config ./configs/coda/csHR2acdcHR_coda.py --checkpoint ./pretrained/CoDA_cs2acdc.pth --out_dir ./workdir/cs2acdc

Traning Steps

python ./tools/train.py --config ./configs/coda/csHR2acdcHR_coda.py --work-dir ./workdir/cs2acdc

chjXu / CoDA