yuhuayc / da-faster-rcnn

An implementation of our CVPR 2018 work 'Domain Adaptive Faster R-CNN for Object Detection in the Wild'

Domain Adaptive Faster R-CNN for Object Detection in the Wild

This is the implementation of our CVPR 2018 work 'Domain Adaptive Faster R-CNN for Object Detection in the Wild'. The aim is to improve the cross-domain robustness of object detection, in the screnario where training and test data are drawn from different distributions. The original paper can be found here.

If you find it helpful for your research, please consider citing:

@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}
}

If you encounter any problems with the code, please contact me at yuhua[dot]chen[at]vision[dot]ee[dot]ethz[dot]ch

Acknowledgment

The implementation is built on the python implementation of Faster RCNN rbgirshick/py-faster-rcnn

Usage

1. Build Caffe and pycaffe (see: Caffe installation instructions)

2. Build the Cython modules

   cd $FRCN_ROOT/lib
   make
   

3. Follow the instrutions of rbgirshick/py-faster-rcnn to download related data.

4. Prepare the dataset, source domain data should start with the filename 'source_', and target domain data with 'target_'.

5. To train the Domain Adaptive Faster R-CNN:

   cd $FRCN_ROOT
   ./tools/train_net.py --gpu {GPU_ID} --solver models/da_faster_rcnn/solver.prototxt --weights data/imagenet_models/VGG16.v2.caffemodel --imdb voc_2007_trainval --iters  {NUM_ITER}  --cfg  {CONFIGURATION_FILE}
   

Example

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

1. Download the datasets from here. Specifically, we will use gtFine_trainvaltest.zip, leftImg8bit_trainvaltest.zip and leftImg8bit_trainvaltest_foggy.zip.

2. Prepare the data using the scripts in 'prepare_data/prepare_data.m'.

3. Train the Domain Adaptive Faster R-CNN:

   cd $FRCN_ROOT
   ./tools/train_net.py --gpu {GPU_ID} --solver models/da_faster_rcnn/solver.prototxt --weights data/imagenet_models/VGG16.v2.caffemodel --imdb voc_2007_trainval --iters  70000  --cfg  models/da_faster_rcnn/faster_rcnn_end2end.yml
   

4. Test the trained model:

   cd $FRCN_ROOT
   ./tools/test_net.py --gpu {GPU_ID} --def models/da_faster_rcnn/test.prototxt --net output/faster_rcnn_end2end/voc_2007_trainval/vgg16_da_faster_rcnn_iter_70000.caffemodel --imdb voc_2007_test --cfg models/da_faster_rcnn/faster_rcnn_end2end.yml
   

Other Implementation

Detectron-DA-Faster-RCNN in Caffe2(Detectron)

Domain-Adaptive-Faster-RCNN-PyTorch in PyTorch