CAUDA

The implementation code of the paper ’Cluster Alignment with Target Knowledge Mining for Unsupervised Domain Adaptation Semantic Segmentation'

Prerequisites

• Create a conda virtual environment and activate it.

conda create --name uncertainty python=3.6
conda activate CAUDA

• Install Pytorch and torchvision following the official instructions, e.g.,

conda install pytorch==1.8.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch

• Install the dependent libraries.

pip install ninja yacs cython matplotlib tqdm opencv-python imageio mmcv

Dataset Preparatation

Download the datasets GTA5 (Download link) and CityScapes (Download link), then place them under the directory like:

semantic-segmentation/datasets
├── gta5
│   ├── images/
│   ├── labels/
│   └── gtav_label_info.p
└── cityscapes
    ├── gtFine/
    └── leftImg8bit/
...

Getting Started

Train

The training code is coming soon.

Evaluation the performances on CityScapes.

• Evaluate the cluster alignment checkpoint

python test.py -cfg configs/deeplabv2_r101_CAUDA.yaml resume results/CAUDA/CAUDA.pth

The checkpoints and logs can be found at link. The extraction code is mimu.

• Evaluate the multi-round self-training checkpoint

python test.py -cfg configs/deeplabv2_r101_tgt_self_distill.yaml resume results/CAUDA_MST/MST.pth

The checkpoints and logs can be found at link. The extraction code is c13y.

Acknowledge

The implementation is based on FADA. Thanks for their great projects.