In this project, we proposed a Domain Disentanglement Faster-RCNN (DDF) for cross-domain object detection, from the view of feature disentanglement.

The implementations are for our paper published in IEEE Transactions on Multimedia:

Decompose to Adapt: Cross-domain Object Detection via Feature Disentanglement

• Python 3+
• Pytorch 1.6.0
• CUDA 11.0

• Cityscapes & Foggy Cityscapes (You can also download the dataset GoogleDrive)
• SIM 10k
• KITTI

We used two pretrained models in our experiments, VGG and ResNet101. You can download these two models from:

• VGG16: Dropbox, VT Server

• ResNet101: Dropbox, VT Server

Download them and put them into the data/pretrained_model/.

You might need to re-build this repository via:

cd lib  
python setup.py build develop

For other detailed settings, please refer to pytorch 1.0 version of repository.

To train the model, please run:

./train_disent.sh

To test the model, please run:

./test_disent.sh

To get the visualization of the feature maps and the feature distance, please refer to:

./test_disent_vis.sh

Please consider citing our papers in your publications if they are helpful to your research:

@article{liu2022decompose,
  title={Decompose to Adapt: Cross-domain Object Detection via Feature Disentanglement},
  author={Liu, Dongnan and Zhang, Chaoyi and Song, Yang and Huang, Heng and Wang, Chenyu and Barnett, Michael and Cai, Weidong},
  journal={IEEE Transactions on Multimedia},
  year={2022},
  publisher={IEEE}
}

Please contact Dongnan Liu (dongnanliu0201@gmail.com) regarding any issues.

DDF is released under the MIT license. See LICENSE for additional details.