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.

Method

• 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.

Results

See presentation and report

Contributors

• Arshad Kazi
• Bhuyashi Deka
• Sadman Sakib