• Our new paper Scale-Aware Domain Adaptive Faster R-CNN has been accepted by IJCV. The corresponding code is maintained under sa-da-faster.

This is a Caffe2 implementation of 'Domain Adaptive Faster R-CNN for Object Detection in the Wild', implemented by Haoran Wang([email protected]). The original paper can be found here. This implementation is built on Detectron @ 5ed75f9.

If you find this repository useful, please cite the oringinal paper:

@inproceedings{chen2018domain,
  title={Domain Adaptive Faster R-CNN for Object Detection in the Wild},
      author =     {Chen, Yuhua and Li, Wen and Sakaridis, Christos and Dai, Dengxin and Van Gool, Luc},
      booktitle =  {Computer Vision and Pattern Recognition (CVPR)},
      year =       {2018}
  }

and Detectron:

@misc{Detectron2018,
  author =       {Ross Girshick and Ilija Radosavovic and Georgia Gkioxari and
                  Piotr Doll\'{a}r and Kaiming He},
  title =        {Detectron},
  howpublished = {\url{https://github.com/facebookresearch/detectron}},
  year =         {2018}
}

Please follow the instruction in Detectron to install and use Detectron-DomainAdaptive-Faster-RCNN.

An example of adapting from Sim10k dataset to Cityscapes dataset is provided:

1. Download the Cityscapes datasets from here and Sim10k datasets from here.

2. Convert the labels of Cityscapes datasets and Sim10k datasets to coco format using the scripts 'tools/convert_cityscapes_to_caronly_coco.py' and 'tools/convert_sim10k_to_coco.py'.

3. Convert ImageNet-pretrained VGG16 Caffe model to Detectron format with 'tools/pickle_caffe_blobs.py' or use my converted VGG16 model in here

4. Train the Domain Adaptive Faster R-CNN:

   cd $DETECTRON
   python2 tools/train_net.py --cfg configs/da_faster_rcnn_baselines/e2e_da_faster_rcnn_vgg16-sim10k.yaml
   

5. Test the trained model:

   cd $DETECTRON
   python2 tools/test_net.py --cfg configs/da_faster_rcnn_baselines/e2e_da_faster_rcnn_vgg16-sim10k.yaml TEST.WEIGHTS /<path_to_trained_model>/model_final.pkl NUM_GPUS 1
   

The best results for different adaptation are reported. Due to the instable nature of adversarial training, the best models are obtained through a model selection on a randomly picked mini validation set.

da-faster-rcnn based on Caffe. (original code by paper authors)

Domain-Adaptive-Faster-RCNN-PyTorch based on PyTorch and maskrcnn-benchmark.

sa-da-faster based on PyTorch and maskrcnn-benchmark.