Unsupervised Domain Adaptation for Semantic Segmentation

We implemented Adaptive Batch Normalization (AdaBN) to adapt an image segmentation model on previously unseen distribution of test data.

Run

To run training, set path and configuration in config.json. Then, set TRAIN=True in main.py and run python3 main.py.

To run inference, set the configuration in infer_config.py. Set, INFER=True in main.py and run python3 main.py.

Ada-BN: Re-calculated Batch norm layer statistics using EMA with multiple pass over test dataset.
Test-Train Mix: Mixed training dataset of different ratios during batchnorm statistics calculation for test dataset.

Name		Name	Last commit message	Last commit date
Latest commit History 13 Commits
results		results
.gitignore		.gitignore
Model_Zoo.py		Model_Zoo.py
Presentation.pdf		Presentation.pdf
README.md		README.md
Report.pdf		Report.pdf
augmentations.py		augmentations.py
config.json		config.json
conventional_dataloader.py		conventional_dataloader.py
dataloader.py		dataloader.py
demo.ipynb		demo.ipynb
infer_config.py		infer_config.py
inference.py		inference.py
losses.py		losses.py
main.py		main.py
metrics.py		metrics.py
requirements.txt		requirements.txt
test.py		test.py
test_augmentations.py		test_augmentations.py
toy_tta.ipynb		toy_tta.ipynb
trainer.py		trainer.py
utils.py		utils.py